Models · Pose estimation

RTMPose

OpenMMLab's RTMPose (t, s, m, l; 256x192 and 384x288): real-time top-down 2D human pose, a CSPNeXt CNN with a SimCC head; also the whole-body DWPose and RTMW checkpoints (133 keypoints with feet, face and hands); Halpe (26 keypoints), hand, face and animal checkpoints. Apache-2.0.

stable Fork of open-mmlab/mmpose@5408bc7 · backends: portable, kernels

Checkpoint

Paper 2023 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 2.90 ms 1080x1920, fp16
vs the faster original 52.0× original fp16: 151 ms
Peak memory 105 MiB original: 459 MiB
Accuracy vs original fp32 +0.0 kernels keypoint AP, COCO val2017 (65.2 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import pose

model = pose.load("rtmpose-t")  # kernels on a supported GPU, else portable PyTorch
result = model(image, boxes=...)
result.keypoints  # also: keypoint_scores, boxes, scores
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench rtmpose-t

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

Input: synthetic, 1080x1920 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call; ≈ 161 ms of it CPU crop + decodefp32 1650.91×475 MiB/524 MiB0.10.17
Original fp16torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call; ≈ 146 ms of it CPU crop + decodefp16 151–459 MiB/490 MiB0.10.15
Original fp32, network onlycrop and decode not timedfp32 4.0238×118 MiB/150 MiB–0.45
Original fp16, network onlycrop and decode not timedfp16 4.8631×89 MiB/114 MiB–0.80
Portable fp32fp32not measured ––––
PortableGPU crop 1.22 · network 2.91 · decode + map back 0.436 msfp16 4.6033×100 MiB/128 MiB0.20.86
KernelsGPU crop 1.24 · network + decode 1.31 · map back 0.325 msfp16 2.9052×105 MiB/156 MiB0.70.29
Portable + torch.compile 1.9179×125 MiB/128 MiB11 (build)–
TensorRT fp16network on person crops onlyfp16 0.866174×56 MiBenginenot comparable16 (build)–
TensorRT + our crop and decodeestimate: engine 0.866 + portable GPU crop 1.22 + decode 0.436 msfp16 2.5260×–––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-10 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. “Network only” times the original’s network on the same crops, batched as one call; the original rows above also include its own crop pipeline and decode, as ours do. The splits under Portable and Kernels come from a separate run that times each step on its own (synchronised in between); the TensorRT engine covers the network only, so its full call is an estimate with our crop and decode added. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-13Originalfp32157 1,804 MiB process–1080x1920
Portablefp1625.6 1,460 MiB processmatches1080x1920
Portablefp3225.7 1,462 MiB processmatches1080x1920
Kernelsfp3220.1 1,532 MiB processmatches1080x1920

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Commercial use allowed

Code
Apache-2.0 LICENSE
Weights
Apache-2.0 source
Training data

Our reading of the licences, not legal advice.

Weights

rtmpose-t-body7-256x192.pth 13.5 MB· hosted on OpenMMLab

fastforks downloads and checks these files on first use (fastforks download rtmpose-t).

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

COCO val2017

ground-truth person boxes, keypoint AP

Rundtypekeypoint AP ↑Δ keypoint AP vs fp32 originalimagesms / image
Original fp32fp3265.2 reference 5000 –
Original fp16fp1665.1 −0.1 5000 –
Portable fp32not measured
Portablefp1665.2 ±0.0 5000 4.56
Kernelsfp1665.2 +0.0 5000 2.08

Top-down on the ground-truth person boxes (non-crowd, at least one labelled keypoint), no flip test; people ranked by the mean of their 17 keypoint scores (not mmpose's rescoring and OKS-NMS), so not the papers' AP; pycocotools OKS AP; whole-body models scored on their 17 body keypoints. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-11; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 2241b095+unc, 2026-11-09.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.66 and p99 <= 3.6 (absolute error of the output keypoints).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 not recorded
GPU default dtypes, mean |diff|≤ 0.66 0.40835
GPU default dtypes, p99 |diff|≤ 3.6 2.02585
gpu_min_score≤ 0.3 not recorded

Recorded for rtmpose-t on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-11, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

RTMPose (Jiang et al. 2023): real-time top-down 2D human pose from OpenMMLab: a CSPNeXt CNN backbone and a SimCC head (a gated attention unit over one token per keypoint, then x and y classified as 1D bins at half-pixel resolution, no heatmaps). The body7 checkpoints (trained on seven public pose datasets) return the 17 COCO keypoints; the DWPose ones (Yang et al. 2023, distilled on COCO-WholeBody + UBody, the pose model of ControlNet's openpose preprocessors) the 133 COCO-WholeBody keypoints (body, feet, face, both hands). The RTMW ones (Jiang et al. 2024: a PAFPN neck and a two-scale head on the same backbone, trained on 14 datasets) return the same 133 keypoints more accurately. More model-zoo checkpoints of the same network: `-halpe26` (body + feet, 26 keypoints), `-hand5` (21 keypoints of a hand box), `-face6` (106 face landmarks of a face box) and `-ap10k` (17 keypoints of an animal box). Takes person boxes (e.g. from `fastforks.detection`) and crops each at 256 x 192 (or 384 x 288; hand, face and animal boxes: 256 x 256); vendored from mmpose (no mmpose / mmcv dependency). The portable `module()` maps crops to SimCC logits.

Input / output
Input: one RGB image and a box per person. Output: keypoints with confidence scores for every person.
Input size
any H x W >= 32, padded to a multiple of 32; tested up to 3000x4000 (453 MiB allocated)
Options
  • flip_test=false· changes the output
Outputs
keypoints, keypoint_scores, boxes, scores
Checkpoint
deepen=0.167, widen=0.375
CUDA kernels
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • fvm_convfused convolutions for image classifiers
  • fvm_extrafused transformer row operations (rotary split, SwiGLU with LayerNorm)
  • fvm_mbfused MBConv blocks (MobileNet and EfficientNet families)
  • fvm_r3further fused classifier blocks
Apple (MPS)
pose_metalMetal: top-down pose decoders and crops. fp32, fp16
Full reference, generated from the repository

OpenMMLab's RTMPose (t, s, m, l; 256x192 and 384x288): real-time top-down 2D human pose, a CSPNeXt CNN with a SimCC head; also the whole-body DWPose and RTMW checkpoints (133 keypoints with feet, face and hands); Halpe (26 keypoints), hand, face and animal checkpoints. Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
rtmpose-tsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
165
151
4.60
2.90
1.91
0.866 (network on person crops only)
475 / 524
459 / 490
100 / 128
105 / 156
125 / 128
127 / 164
rtmpose-ssynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
161
165
4.95
3.08
2.07
1.06 (network on person crops only)
461 / 510
449 / 480
118 / 152
109 / 158
129 / 150
137 / 170
rtmpose-msynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
179
180
6.36
4.50
3.42
2.02 (network on person crops only)
323 / 376
316 / 364
145 / 192
125 / 200
147 / 172
157 / 176
rtmpose-lsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
161
176
9.04
5.70
5.32
3.20 (network on person crops only)
1076 / 1114
1105 / 1154
200 / 256
152 / 244
185 / 194
230 / 266
rtmpose-m-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
174
167
13.3
9.23
7.09
4.33 (network on person crops only)
585 / 666
505 / 582
224 / 332
178 / 320
190 / 212
212 / 242
rtmpose-l-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
192
171
19.4
11.4
10.7
7.08 (network on person crops only)
1485 / 1538
366 / 410
314 / 416
217 / 390
242 / 274
297 / 322
dwpose-tsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
160
174
4.60
3.17
1.96
0.940 (network on person crops only)
841 / 874
832 / 858
104 / 132
109 / 174
129 / 148
136 / 160
dwpose-ssynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
173
180
5.16
3.35
2.27
1.17 (network on person crops only)
807 / 860
806 / 848
124 / 154
115 / 178
135 / 150
149 / 170
dwpose-msynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
169
164
7.31
4.86
3.80
2.13 (network on person crops only)
746 / 802
756 / 792
153 / 196
133 / 200
156 / 174
203 / 228
dwpose-lsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
178
184
10.3
6.20
5.75
3.33 (network on person crops only)
341 / 388
393 / 428
211 / 260
163 / 266
197 / 216
277 / 294
dwpose-l-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
194
203
21.4
12.7
11.9
7.46 (network on person crops only)
796 / 832
773 / 816
325 / 418
229 / 404
253 / 296
347 / 370
rtmw-msynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
169
205
10.9
7.54
6.23
3.34 (network on person crops only)
288 / 328
379 / 410
181 / 244
164 / 266
183 / 202
231 / 256
rtmw-lsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
213
203
15.3
9.99
9.34
5.17 (network on person crops only)
679 / 708
865 / 908
256 / 308
217 / 312
242 / 264
354 / 396
rtmw-xsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
218
207
24.0
14.4
15.3
8.47 (network on person crops only)
1307 / 1360
604 / 636
335 / 390
289 / 392
320 / 328
497 / 514
rtmw-l-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
274
275
31.8
20.6
19.5
11.2 (network on person crops only)
982 / 1018
1094 / 1144
371 / 468
276 / 458
298 / 344
424 / 452
rtmw-x-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
306
291
50.6
50.9
34.0
19.3 (network on person crops only)
1705 / 1812
1874 / 1940
466 / 632
496 / 632
361 / 388
581 / 624
rtmpose-t-halpe26synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
181
172
4.61
2.90
1.94
0.871 (network on person crops only)
119 / 158
103 / 144
100 / 130
105 / 156
125 / 128
128 / 164
rtmpose-s-halpe26synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
172
173
4.99
3.05
2.11
1.07 (network on person crops only)
174 / 212
162 / 200
119 / 152
110 / 178
130 / 150
138 / 170
rtmpose-m-halpe26synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
179
181
6.42
4.53
3.44
2.02 (network on person crops only)
288 / 328
283 / 318
146 / 192
125 / 180
148 / 172
160 / 180
rtmpose-l-halpe26synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
180
164
9.10
5.69
5.34
3.20 (network on person crops only)
498 / 532
529 / 572
201 / 256
152 / 244
186 / 194
231 / 266
rtmpose-m-halpe26-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
192
193
13.4
9.28
7.12
4.34 (network on person crops only)
769 / 850
690 / 764
225 / 332
179 / 320
190 / 212
212 / 238
rtmpose-l-halpe26-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
206
206
19.5
11.5
10.6
7.08 (network on person crops only)
1019 / 1054
975 / 1038
315 / 416
218 / 410
243 / 276
299 / 322
rtmpose-x-halpe26-384synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
247
217
33.8
19.1
21.1
13.3 (network on person crops only)
1470 / 1588
470 / 556
385 / 478
280 / 448
262 / 298
430 / 456
rtmpose-m-hand5synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
192
195
8.05
5.46
4.51
2.58 (network on person crops only)
451 / 504
425 / 472
166 / 214
139 / 200
176 / 192
174 / 192
rtmpose-t-face6synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
188
204
4.53
3.61
2.23
1.00 (network on person crops only)
444 / 498
424 / 464
116 / 152
123 / 180
159 / 168
145 / 168
rtmpose-s-face6synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
258
237
5.04
3.74
2.69
1.28 (network on person crops only)
519 / 572
504 / 540
140 / 174
128 / 202
164 / 190
158 / 178
rtmpose-m-face6synthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
198
207
8.75
5.77
4.79
2.67 (network on person crops only)
666 / 728
652 / 698
173 / 216
145 / 222
182 / 192
216 / 240
rtmpose-m-ap10ksynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
196
186
8.02
5.42
4.54
2.57 (network on person crops only)
783 / 846
757 / 814
166 / 214
139 / 200
176 / 192
172 / 188
Details
  • Original code: https://github.com/open-mmlab/mmpose at 5408bc76f5b8
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
  • Licence evidence: https://github.com/open-mmlab/mmpose/blob/5408bc76f5b848cf925a0d1857899011d8c5b497/README.md#license
  • Training data: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7: COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18 keypoints; the CSPNeXt backbone was pre-trained on ImageNet-1k (and on the same pose data, pt-body7))
    • dwpose-t: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), ImageNet-1k (ILSVRC 2012 classification) (COCO-WholeBody and UBody whole-body keypoints, distilled from a larger RTMPose (DWPose two-stage distillation); the CSPNeXt backbone was pre-trained on ImageNet-1k and on AI Challenger + COCO body keypoints)
    • dwpose-s: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), ImageNet-1k (ILSVRC 2012 classification) (COCO-WholeBody and UBody whole-body keypoints, distilled from a larger RTMPose (DWPose two-stage distillation); the CSPNeXt backbone was pre-trained on ImageNet-1k and on AI Challenger + COCO body keypoints)
    • dwpose-m: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), ImageNet-1k (ILSVRC 2012 classification) (COCO-WholeBody and UBody whole-body keypoints, distilled from a larger RTMPose (DWPose two-stage distillation); the CSPNeXt backbone was pre-trained on ImageNet-1k and on AI Challenger + COCO body keypoints)
    • dwpose-l: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), ImageNet-1k (ILSVRC 2012 classification) (COCO-WholeBody and UBody whole-body keypoints, distilled from a larger RTMPose (DWPose two-stage distillation); the CSPNeXt backbone was pre-trained on ImageNet-1k and on AI Challenger + COCO body keypoints)
    • dwpose-l-384: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), ImageNet-1k (ILSVRC 2012 classification) (COCO-WholeBody and UBody whole-body keypoints, distilled from a larger RTMPose (DWPose two-stage distillation); the CSPNeXt backbone was pre-trained on ImageNet-1k and on AI Challenger + COCO body keypoints)
    • rtmw-m: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, COCO 2017 (train, instance annotations), ImageNet-1k (ILSVRC 2012 classification) (Cocktail14: AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe, PoseTrack18, COCO-WholeBody, UBody, Human-Art, WFLW, 300W, COFW, LaPa and InterHand2.6M (the last six are not in the dataset list; several are research-only), distilled from RTMW-l; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body / whole-body keypoints)
    • rtmw-l: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, COCO 2017 (train, instance annotations), ImageNet-1k (ILSVRC 2012 classification) (Cocktail14: AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe, PoseTrack18, COCO-WholeBody, UBody, Human-Art, WFLW, 300W, COFW, LaPa and InterHand2.6M (the last six are not in the dataset list; several are research-only), distilled from RTMW-x; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body / whole-body keypoints)
    • rtmw-x: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, COCO 2017 (train, instance annotations), ImageNet-1k (ILSVRC 2012 classification) (Cocktail14: AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe, PoseTrack18, COCO-WholeBody, UBody, Human-Art, WFLW, 300W, COFW, LaPa and InterHand2.6M (the last six are not in the dataset list; several are research-only); the CSPNeXt backbone was pre-trained on ImageNet-1k and on body / whole-body keypoints)
    • rtmw-l-384: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, COCO 2017 (train, instance annotations), ImageNet-1k (ILSVRC 2012 classification) (Cocktail14: AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe, PoseTrack18, COCO-WholeBody, UBody, Human-Art, WFLW, 300W, COFW, LaPa and InterHand2.6M (the last six are not in the dataset list; several are research-only), distilled from RTMW-x; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body / whole-body keypoints)
    • rtmw-x-384: COCO-WholeBody, UBody (upper-body whole-body keypoints, from OSX), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, COCO 2017 (train, instance annotations), ImageNet-1k (ILSVRC 2012 classification) (Cocktail14: AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe, PoseTrack18, COCO-WholeBody, UBody, Human-Art, WFLW, 300W, COFW, LaPa and InterHand2.6M (the last six are not in the dataset list; several are research-only); the CSPNeXt backbone was pre-trained on ImageNet-1k and on body / whole-body keypoints)
    • rtmpose-t-halpe26: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7 with the 26 Halpe keypoints (COCO-WholeBody feet annotations for COCO): COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body7)
    • rtmpose-s-halpe26: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7 with the 26 Halpe keypoints (COCO-WholeBody feet annotations for COCO): COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body7)
    • rtmpose-m-halpe26: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7 with the 26 Halpe keypoints (COCO-WholeBody feet annotations for COCO): COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body7)
    • rtmpose-l-halpe26: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7 with the 26 Halpe keypoints (COCO-WholeBody feet annotations for COCO): COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body7)
    • rtmpose-m-halpe26-384: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7 with the 26 Halpe keypoints (COCO-WholeBody feet annotations for COCO): COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body7)
    • rtmpose-l-halpe26-384: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7 with the 26 Halpe keypoints (COCO-WholeBody feet annotations for COCO): COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body7)
    • rtmpose-x-halpe26-384: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018, ImageNet-1k (ILSVRC 2012 classification) (body7 with the 26 Halpe keypoints (COCO-WholeBody feet annotations for COCO): COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18; the CSPNeXt backbone was pre-trained on ImageNet-1k and on body7)
    • rtmpose-m-hand5: COCO-WholeBody, Halpe-FullBody, COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), ImageNet-1k (ILSVRC 2012 classification) (hand5: COCO-WholeBody-Hand, OneHand10K, FreiHAND, RHD and Halpe hands (OneHand10K, FreiHAND and RHD are not in the dataset list; research-only terms); the CSPNeXt backbone was pre-trained on ImageNet-1k and on AI Challenger + COCO body keypoints)
    • rtmpose-t-face6: COCO-WholeBody, Halpe-FullBody, ImageNet-1k (ILSVRC 2012 classification) (face6: COCO-WholeBody-Face, WFLW, 300W, COFW, Halpe faces and LaPa, converted to the 106 LaPa landmarks (WFLW, 300W, COFW and LaPa are not in the dataset list; research-only terms); the CSPNeXt backbone was pre-trained on ImageNet-1k)
    • rtmpose-s-face6: COCO-WholeBody, Halpe-FullBody, ImageNet-1k (ILSVRC 2012 classification) (face6: COCO-WholeBody-Face, WFLW, 300W, COFW, Halpe faces and LaPa, converted to the 106 LaPa landmarks (WFLW, 300W, COFW and LaPa are not in the dataset list; research-only terms); the CSPNeXt backbone was pre-trained on ImageNet-1k)
    • rtmpose-m-face6: COCO-WholeBody, Halpe-FullBody, ImageNet-1k (ILSVRC 2012 classification) (face6: COCO-WholeBody-Face, WFLW, 300W, COFW, Halpe faces and LaPa, converted to the 106 LaPa landmarks (WFLW, 300W, COFW and LaPa are not in the dataset list; research-only terms); the CSPNeXt backbone was pre-trained on ImageNet-1k)
    • rtmpose-m-ap10k: AP-10K (animal pose), COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), ImageNet-1k (ILSVRC 2012 classification) (AP-10K (animal keypoints, 54 species); the CSPNeXt backbone was pre-trained on ImageNet-1k and on AI Challenger + COCO body keypoints)
  • Options: flip_test = False
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: rtmpose-t, rtmpose-s, rtmpose-m, rtmpose-l, rtmpose-m-384, rtmpose-l-384, dwpose-t, dwpose-s, dwpose-m, dwpose-l, dwpose-l-384, rtmw-m, rtmw-l, rtmw-x, rtmw-l-384, rtmw-x-384, rtmpose-t-halpe26, rtmpose-s-halpe26, rtmpose-m-halpe26, rtmpose-l-halpe26, rtmpose-m-halpe26-384, rtmpose-l-halpe26-384, rtmpose-x-halpe26-384, rtmpose-m-hand5, rtmpose-t-face6, rtmpose-s-face6, rtmpose-m-face6, rtmpose-m-ap10k
  • Outputs: keypoints, keypoint_scores, boxes, scores; label_set: coco-keypoints-17
  • Input size: any H x W >= 32, padded to a multiple of 32
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): rtmpose-t 3000x4000 (0.4 GiB peak allocated, 0.5 reserved); rtmpose-s 3000x4000 (0.4 GiB peak allocated, 0.5 reserved); rtmpose-m 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmpose-l 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmpose-m-384 3000x4000 (0.5 GiB peak allocated, 0.6 reserved); rtmpose-l-384 3000x4000 (0.6 GiB peak allocated, 1.0 reserved); dwpose-t 3000x4000 (0.4 GiB peak allocated, 0.5 reserved); dwpose-s 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); dwpose-m 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); dwpose-l 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); dwpose-l-384 3000x4000 (0.6 GiB peak allocated, 1.0 reserved); rtmw-m 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmw-l 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmw-x 3000x4000 (0.6 GiB peak allocated, 0.7 reserved); rtmw-l-384 3000x4000 (0.7 GiB peak allocated, 1.0 reserved); rtmw-x-384 3000x4000 (0.8 GiB peak allocated, 1.0 reserved); rtmpose-t-halpe26 3000x4000 (0.4 GiB peak allocated, 0.5 reserved); rtmpose-s-halpe26 3000x4000 (0.4 GiB peak allocated, 0.5 reserved); rtmpose-m-halpe26 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmpose-l-halpe26 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmpose-m-halpe26-384 3000x4000 (0.5 GiB peak allocated, 0.6 reserved); rtmpose-l-halpe26-384 3000x4000 (0.6 GiB peak allocated, 1.0 reserved); rtmpose-x-halpe26-384 3000x4000 (0.8 GiB peak allocated, 0.9 reserved); rtmpose-m-hand5 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmpose-t-face6 3000x4000 (0.4 GiB peak allocated, 0.5 reserved); rtmpose-s-face6 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmpose-m-face6 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmpose-m-ap10k 3000x4000 (0.5 GiB peak allocated, 0.5 reserved)
  • rtmpose-t 1080x1920, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 16 s (ONNX + engine).
  • rtmpose-s 1080x1920, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 16 s (ONNX + engine).
  • rtmpose-m 1080x1920, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 18 s (ONNX + engine).
  • rtmpose-l 1080x1920, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 19 s (ONNX + engine).
  • rtmpose-m-384 1080x1920, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 19 s (ONNX + engine).
  • rtmpose-l-384 1080x1920, first call: portable + torch.compile 15 s (cold Inductor cache), TensorRT fp16 20 s (ONNX + engine).
  • dwpose-t 1080x1920, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 16 s (ONNX + engine).
  • dwpose-s 1080x1920, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 17 s (ONNX + engine).
  • dwpose-m 1080x1920, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 19 s (ONNX + engine).
  • dwpose-l 1080x1920, first call: portable + torch.compile 15 s (cold Inductor cache), TensorRT fp16 20 s (ONNX + engine).
  • dwpose-l-384 1080x1920, first call: portable + torch.compile 15 s (cold Inductor cache), TensorRT fp16 21 s (ONNX + engine).
  • rtmw-m 1080x1920, first call: portable + torch.compile 19 s (cold Inductor cache), TensorRT fp16 23 s (ONNX + engine).
  • rtmw-l 1080x1920, first call: portable + torch.compile 19 s (cold Inductor cache), TensorRT fp16 25 s (ONNX + engine).
  • rtmw-x 1080x1920, first call: portable + torch.compile 20 s (cold Inductor cache), TensorRT fp16 27 s (ONNX + engine).
  • rtmw-l-384 1080x1920, first call: portable + torch.compile 20 s (cold Inductor cache), TensorRT fp16 27 s (ONNX + engine).
  • rtmw-x-384 1080x1920, first call: portable + torch.compile 21 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).
  • rtmpose-t-halpe26 1080x1920, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 17 s (ONNX + engine).
  • rtmpose-s-halpe26 1080x1920, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 17 s (ONNX + engine).
  • rtmpose-m-halpe26 1080x1920, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 18 s (ONNX + engine).
  • rtmpose-l-halpe26 1080x1920, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 20 s (ONNX + engine).
  • rtmpose-m-halpe26-384 1080x1920, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 19 s (ONNX + engine).
  • rtmpose-l-halpe26-384 1080x1920, first call: portable + torch.compile 15 s (cold Inductor cache), TensorRT fp16 21 s (ONNX + engine).
  • rtmpose-x-halpe26-384 1080x1920, first call: portable + torch.compile 17 s (cold Inductor cache), TensorRT fp16 23 s (ONNX + engine).
  • rtmpose-m-hand5 1080x1920, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 18 s (ONNX + engine).
  • rtmpose-t-face6 1080x1920, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 16 s (ONNX + engine).
  • rtmpose-s-face6 1080x1920, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 17 s (ONNX + engine).
  • rtmpose-m-face6 1080x1920, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 18 s (ONNX + engine).
  • rtmpose-m-ap10k 1080x1920, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 18 s (ONNX + engine).
  • rtmpose-t original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-t original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-s original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-s original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-384 original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l-384 original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-t original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-t original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-s original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-s original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-m original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-m original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-l original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-l original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-l-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • dwpose-l-384 original 1080x1920: torch.autocast(fp16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-m original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-m original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-l original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-l original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-x original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-x original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-l-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-l-384 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-x-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmw-x-384 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-t-halpe26 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-t-halpe26 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-s-halpe26 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-s-halpe26 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-halpe26 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-halpe26 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l-halpe26 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l-halpe26 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-halpe26-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-halpe26-384 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l-halpe26-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-l-halpe26-384 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-x-halpe26-384 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-x-halpe26-384 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-hand5 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-hand5 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-t-face6 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-t-face6 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-s-face6 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-s-face6 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-face6 original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-face6 original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-ap10k original (fp32) 1080x1920: mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-m-ap10k original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_topdown (CPU cv2 crop, network, SimCC decode; flip_test off), all in the timed call
  • rtmpose-t load / first call: original 0.13 s / 0.152 s, portable 0.1 s / 0.879 s, kernels 1.42 s / 0.291 s
  • rtmpose-s load / first call: original 0.16 s / 0.164 s, portable 0.1 s / 1.033 s, kernels 1.4 s / 0.291 s
  • rtmpose-m load / first call: original 0.23 s / 0.156 s, portable 0.22 s / 1.105 s, kernels 1.5 s / 0.291 s
  • rtmpose-l load / first call: original 0.38 s / 0.179 s, portable 0.94 s / 0.858 s, kernels 1.6 s / 0.29 s
  • rtmpose-m-384 load / first call: original 0.23 s / 0.349 s, portable 0.2 s / 1.245 s, kernels 1.48 s / 0.29 s
  • rtmpose-l-384 load / first call: original 0.58 s / 0.165 s, portable 0.41 s / 1.0 s, kernels 1.56 s / 0.291 s
  • dwpose-t load / first call: original 0.16 s / 0.326 s, portable 0.17 s / 1.243 s, kernels 0.69 s / 0.29 s
  • dwpose-s load / first call: original 0.19 s / 0.47 s, portable 0.12 s / 1.022 s, kernels 1.48 s / 0.29 s
  • dwpose-m load / first call: original 0.26 s / 0.596 s, portable 0.24 s / 1.099 s, kernels 1.51 s / 0.292 s
  • dwpose-l load / first call: original 0.5 s / 1.143 s, portable 0.88 s / 1.831 s, kernels 1.64 s / 0.28 s
  • dwpose-l-384 load / first call: original 0.42 s / 0.425 s, portable 0.58 s / 1.102 s, kernels 1.67 s / 0.285 s
  • rtmw-m load / first call: original 0.54 s / 1.213 s, portable 0.32 s / 0.896 s, kernels 1.02 s / 0.301 s
  • rtmw-l load / first call: original 0.76 s / 0.632 s, portable 0.39 s / 1.032 s, kernels 1.72 s / 0.306 s
  • rtmw-x load / first call: original 1.23 s / 0.64 s, portable 1.35 s / 0.833 s, kernels 2.1 s / 0.283 s
  • rtmw-l-384 load / first call: original 0.75 s / 0.56 s, portable 0.34 s / 1.198 s, kernels 1.69 s / 0.279 s
  • rtmw-x-384 load / first call: original 1.05 s / 0.419 s, portable 0.92 s / 0.97 s, kernels 1.36 s / 0.329 s
  • rtmpose-t-halpe26 load / first call: original 0.18 s / 0.707 s, portable 0.16 s / 0.872 s, kernels 0.73 s / 0.292 s
  • rtmpose-s-halpe26 load / first call: original 0.19 s / 0.614 s, portable 0.07 s / 0.943 s, kernels 1.46 s / 0.296 s
  • rtmpose-m-halpe26 load / first call: original 0.25 s / 0.445 s, portable 0.14 s / 1.037 s, kernels 1.51 s / 0.297 s
  • rtmpose-l-halpe26 load / first call: original 0.37 s / 0.451 s, portable 0.25 s / 1.128 s, kernels 1.55 s / 0.292 s
  • rtmpose-m-halpe26-384 load / first call: original 0.26 s / 0.341 s, portable 0.13 s / 1.15 s, kernels 1.48 s / 0.292 s
  • rtmpose-l-halpe26-384 load / first call: original 0.41 s / 0.382 s, portable 0.26 s / 1.239 s, kernels 1.57 s / 0.29 s
  • rtmpose-x-halpe26-384 load / first call: original 0.89 s / 0.938 s, portable 0.41 s / 1.17 s, kernels 1.76 s / 0.276 s
  • rtmpose-m-hand5 load / first call: original 0.23 s / 0.219 s, portable 0.13 s / 1.134 s, kernels 1.49 s / 0.289 s
  • rtmpose-t-face6 load / first call: original 0.18 s / 0.339 s, portable 0.16 s / 1.295 s, kernels 0.71 s / 0.292 s
  • rtmpose-s-face6 load / first call: original 0.31 s / 0.354 s, portable 0.17 s / 1.11 s, kernels 0.69 s / 0.296 s
  • rtmpose-m-face6 load / first call: original 0.29 s / 0.203 s, portable 0.15 s / 1.146 s, kernels 1.49 s / 0.289 s
  • rtmpose-m-ap10k load / first call: original 0.23 s / 0.19 s, portable 0.13 s / 1.139 s, kernels 1.48 s / 0.291 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Top-down on the same every-tenth-image subset for all modes, ground-truth person boxes, no flip test; people ranked by the mean of their first 17 keypoint scores; pycocotools OKS AP on COCO body joints. Top-down on the ground-truth person boxes (non-crowd, at least one labelled keypoint), no flip test; people ranked by the mean of their 17 keypoint scores (not mmpose's rescoring and OKS-NMS), so not the papers' AP; pycocotools OKS AP; whole-body models scored on their 17 body keypoints. One prediction per annotated hand box. PCK is the mean across landmarks of the fraction within 20% of the longer box side. EPE is the mean pixel distance across annotated landmarks. Flip testing is disabled. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
rtmpose-tCOCO val2017, ground-truth person boxes, keypoint AP65.265.165.265.2
rtmpose-sCOCO val2017, ground-truth person boxes, keypoint AP69.869.869.869.8
rtmpose-mCOCO val2017, ground-truth person boxes, keypoint AP75.675.675.675.6
rtmpose-lCOCO val2017, ground-truth person boxes, keypoint AP77.977.977.978.0
rtmpose-m-384COCO val2017, ground-truth person boxes, keypoint AP78.077.977.978.0
rtmpose-l-384COCO val2017, ground-truth person boxes, keypoint AP80.080.080.079.9
dwpose-tCOCO val2017, ground-truth person boxes, keypoint AP60.260.260.160.1
dwpose-sCOCO val2017, ground-truth person boxes, keypoint AP65.665.665.665.5
dwpose-mCOCO val2017, ground-truth person boxes, keypoint AP71.771.871.871.8
dwpose-lCOCO val2017, ground-truth person boxes, keypoint AP73.973.974.074.0
dwpose-l-384COCO val2017, ground-truth person boxes, keypoint AP75.675.675.775.6
rtmw-mCOCO val2017, ground-truth person boxes, keypoint AP74.574.574.574.6
rtmw-lCOCO val2017, ground-truth person boxes, keypoint AP76.976.977.076.9
rtmw-xCOCO val2017, ground-truth person boxes, keypoint AP77.377.477.377.4
rtmw-l-384COCO val2017, every 10th image, GT person boxes, keypoint AP80.280.380.280.180.1
rtmw-l-384COCO val2017, ground-truth person boxes, keypoint AP79.479.479.579.5
rtmw-x-384COCO val2017, every 10th image, GT person boxes, keypoint AP80.380.280.180.380.4
rtmw-x-384COCO val2017, ground-truth person boxes, keypoint AP79.679.679.679.6
rtmpose-t-halpe26COCO val2017, ground-truth person boxes, keypoint AP65.365.565.365.3
rtmpose-s-halpe26COCO val2017, every 10th image, GT person boxes, keypoint AP71.471.371.471.2
rtmpose-s-halpe26COCO val2017, ground-truth person boxes, keypoint AP70.070.069.970.0
rtmpose-m-halpe26COCO val2017, ground-truth person boxes, keypoint AP76.376.276.276.2
rtmpose-l-halpe26COCO val2017, ground-truth person boxes, keypoint AP78.578.578.578.5
rtmpose-m-halpe26-384COCO val2017, ground-truth person boxes, keypoint AP78.378.478.378.3
rtmpose-l-halpe26-384COCO val2017, ground-truth person boxes, keypoint AP79.879.879.880.0
rtmpose-x-halpe26-384COCO val2017, ground-truth person boxes, keypoint AP80.580.680.580.4
rtmpose-m-hand5COCO-WholeBody validation (1,568 images), hand [email protected] % / EPE in px80.98 / 4.56

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 2241b095+uncommitted / be54b8b0, 2026-11-09 to 2026-11-11; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 3bf0bac7 / 44640884, 2026-11-18; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 2241b095+uncommitted / be54b8b0, 2026-11-09 to 2026-11-11.

Parity

CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.66 and p99 <= 3.6 (absolute error of the output keypoints).

variantkernels error vs original, mean / p99GPU, torchkernels used
rtmpose-t0.40835 / 2.02585RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-s0.35962 / 2.04983RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-m0.14799 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-l0.19814 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-m-3840.32639 / 2.09004RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-l-3840.2774 / 0.79916RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
dwpose-t0.18757 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
dwpose-s0.21925 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
dwpose-m0.09716 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
dwpose-l0.16913 / 1.64226RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
dwpose-l-3840.16432 / 2.22965RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmw-m0.10568 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmw-l0.07894 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmw-x0.12505 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmw-l-3840.94198 / 1.29593RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmw-x-3840.68362 / 2.91205RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-t-halpe260.51209 / 4.77095RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-s-halpe261.41323 / 27.36431RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-m-halpe260.24452 / 1.78611RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-l-halpe260.20565 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-m-halpe26-3840.16174 / 1.48645RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-l-halpe26-3840.14576 / 0.79916RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-x-halpe26-3840.15131 / 0.79918RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-m-hand50.39958 / 2.61323RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-t-face60.12106 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-s-face60.10929 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-m-face60.07662 / 1.19873RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
rtmpose-m-ap10k0.52728 / 4.73499RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input
Output
Astronaut Eileen Collins, input
Input · 100%

Input: Astronaut Eileen Collins, NASA (public domain).