Models · Pose estimation

RTMO

OpenMMLab's RTMO (t, s, m, l): real-time one-stage multi-person 2D pose; finds people and their 17 COCO keypoints in one pass over the image, no person detector needed. 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 6.31 ms 1080x1920, fp16
vs the faster original 1.8× original fp32: 11.3 ms
Peak memory 121 MiB original: 90 MiB
Accuracy vs original fp32 −0.0 kernels keypoint AP, COCO val2017 (57.4 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("rtmo-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 rtmo-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 fp32Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed callfp32 11.3–90 MiB/108 MiB5.30.43
Original fp16torch.autocast(float16) around the call; mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed callfp16 13.30.85×118 MiB/122 MiB0.30.32
Portable fp32fp32not measured ––––
Portablefp16 7.621.5×121 MiB/136 MiB0.10.41
Kernelsfp16 6.311.8×121 MiB/156 MiB1.40.28
Portable + torch.compile 3.303.4×111 MiB/134 MiB16 (build)–
TensorRT fp16network on person crops onlyfp16 0.88613×40 MiBenginenot comparable37 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-11. 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. 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, 2026-11-12Originalfp3225.4 1,836 MiB process–1080x1920
Portablefp1619.9 1,506 MiB processmatches1080x1920
Portablefp3221.0 1,523 MiB processmatches1080x1920
Kernelsfp16not implemented–1080x1920

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

rtmo-t-body7-416x416.pth 27.6 MB· hosted on OpenMMLab

fastforks downloads and checks these files on first use (fastforks download rtmo-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

predicted people, keypoint AP

Rundtypekeypoint AP ↑Δ keypoint AP vs fp32 originalimagesms / image
Original fp32fp3257.4 reference 5000 –
Original fp16fp1657.4 ±0.0 5000 –
Portable fp32not measured
Portablefp1657.4 −0.0 5000 –
Kernelsfp1657.4 −0.0 5000 –

Bottom-up on all 5,000 validation images, no ground-truth boxes; people ranked by their predicted person score, after model-specific decoding; pycocotools OKS AP. 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 be54b8b0, 2026-11-11.

COCO val2017

every 10th image, predicted people, keypoint AP

Rundtypekeypoint AP ↑Δ keypoint AP vs fp32 originalimagesms / image
Original fp32fp3259.4 reference 500 –
Original fp16fp1659.4 ±0.0 500 –
Portable fp32fp3259.3 −0.0 500 9.31
Portablefp1659.3 −0.1 500 –
Kernelsfp1659.3 −0.0 500 –

Bottom-up on the same every-tenth-image subset for all modes, ranked by predicted person score after model-specific decoding; pycocotools OKS AP. 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 fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 44640884, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks be54b8b0, 2026-11-11.

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| <= 0.014 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.47 and p99 <= 9.8 (absolute error of the output keypoints).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.014 not recorded
GPU default dtypes, mean |diff|≤ 0.47 0.02725
GPU default dtypes, p99 |diff|≤ 9.8 0.14195
gpu_min_score≤ 0.3 not recorded

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

Details

RTMO (Lu et al. 2024): real-time one-stage multi-person 2D pose from OpenMMLab. A YOLOX-style CSPDarknet backbone, a hybrid encoder neck (one transformer layer on the coarsest map) and a head that predicts, per grid cell, a person score, a box and a pose vector; a dynamic coordinate classifier turns each kept pose vector into the 17 COCO keypoints over 1D bins that span the person's box. Takes the whole image without person boxes or a separate detector: the backbone runs once per image, while decoding and NMS depend on the detections. The image is letterboxed to 640 x 640 (416 x 416 for `rtmo-t`); vendored from mmpose (no mmpose / mmcv dependency). The portable `module()` maps the letterboxed image to the head's per-cell outputs (decoding and NMS stay outside).

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 (460 MiB allocated)
Options
  • score_threshold=0.10.0 to 1.0· changes the output
Outputs
keypoints, keypoint_scores, boxes, scores
Checkpoint
deepen=0.33, widen=0.375, size=416
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)
fp32, fp16

Limitations

  • module() exports the letterboxed-image network and raw per-cell head outputs; coordinate decoding and NMS remain outside the exported graph.
  • Float16 and float32 can retain different people near the score threshold.
Full reference, generated from the repository

OpenMMLab's RTMO (t, s, m, l): real-time one-stage multi-person 2D pose; finds people and their 17 COCO keypoints in one pass over the image, no person detector needed. Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
rtmo-tsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
11.3
13.3
7.62
6.31
3.30
0.886 (network on person crops only)
90 / 108
118 / 122
121 / 136
121 / 156
111 / 134
112 / 150
rtmo-ssynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
11.4
15.5
7.48
6.27
3.35
1.21 (network on person crops only)
192 / 220
228 / 252
142 / 160
142 / 180
121 / 156
133 / 164
rtmo-msynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
14.3
17.0
8.71
6.85
4.99
2.03 (network on person crops only)
332 / 360
422 / 462
166 / 190
166 / 212
146 / 170
194 / 232
rtmo-lsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
17.1
21.4
10.8
8.30
7.07
2.90 (network on person crops only)
614 / 652
800 / 824
208 / 226
208 / 228
192 / 224
277 / 306
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 (body7: COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18 keypoints; the CSPDarknet backbone starts from a YOLOX COCO detection checkpoint)
  • Options: score_threshold = 0.1 (0.0 to 1.0)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: rtmo-t, rtmo-s, rtmo-m, rtmo-l
  • 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): rtmo-t 3000x4000 (0.4 GiB peak allocated, 0.5 reserved); rtmo-s 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmo-m 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmo-l 3000x4000 (0.5 GiB peak allocated, 0.6 reserved)
  • rtmo-t 1080x1920, first call: portable + torch.compile 16 s (cold Inductor cache), TensorRT fp16 37 s (ONNX + engine).
  • rtmo-s 1080x1920, first call: portable + torch.compile 15 s (cold Inductor cache), TensorRT fp16 37 s (ONNX + engine).
  • rtmo-m 1080x1920, first call: portable + torch.compile 18 s (cold Inductor cache), TensorRT fp16 39 s (ONNX + engine).
  • rtmo-l 1080x1920, first call: portable + torch.compile 19 s (cold Inductor cache), TensorRT fp16 40 s (ONNX + engine).
  • rtmo-t original (fp32) 1080x1920: mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-t original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-s original (fp32) 1080x1920: mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-s original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-m original (fp32) 1080x1920: mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-m original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-l original (fp32) 1080x1920: mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-l original 1080x1920: torch.autocast(float16) around the call; mmpose 1.3.2 inference_bottomup (CPU cv2 letterbox, network, NMS, decode), all in the timed call
  • rtmo-t load / first call: original 0.26 s / 0.324 s, portable 0.11 s / 0.411 s, kernels 1.38 s / 0.279 s
  • rtmo-s load / first call: original 0.27 s / 0.076 s, portable 0.12 s / 0.412 s, kernels 1.37 s / 0.264 s
  • rtmo-m load / first call: original 0.37 s / 0.077 s, portable 0.22 s / 0.428 s, kernels 1.43 s / 0.242 s
  • rtmo-l load / first call: original 0.55 s / 0.063 s, portable 0.36 s / 0.243 s, kernels 1.52 s / 0.16 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Bottom-up on all 5,000 validation images, no ground-truth boxes; people ranked by their predicted person score, after model-specific decoding; pycocotools OKS AP. Bottom-up on the same every-tenth-image subset for all modes, ranked by predicted person score after model-specific decoding; pycocotools OKS AP. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
rtmo-tCOCO val2017, predicted people, keypoint AP57.457.457.457.4
rtmo-tCOCO val2017, every 10th image, predicted people, keypoint AP59.459.459.359.359.3
rtmo-sCOCO val2017, predicted people, keypoint AP68.668.668.568.5
rtmo-sCOCO val2017, every 10th image, predicted people, keypoint AP71.171.271.271.271.2
rtmo-mCOCO val2017, predicted people, keypoint AP72.672.672.672.6
rtmo-mCOCO val2017, every 10th image, predicted people, keypoint AP74.274.274.274.274.2
rtmo-lCOCO val2017, predicted people, keypoint AP74.874.874.874.8
rtmo-lCOCO val2017, every 10th image, predicted people, keypoint AP76.176.076.076.176.0

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks be54b8b0, 2026-11-11; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 44640884, 2026-11-17 to 2026-11-18; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks be54b8b0, 2026-11-11.

Parity

CPU, float32: max |diff| <= 0.014 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.47 and p99 <= 9.8 (absolute error of the output keypoints).

variantkernels error vs original, mean / p99GPU, torchkernels used
rtmo-t0.02725 / 0.14195RTX 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
rtmo-s0.02234 / 0.14143RTX 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
rtmo-m0.11064 / 1.21749RTX 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
rtmo-l0.30973 / 6.5301RTX 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

Limitations

  • module() exports the letterboxed-image network and raw per-cell head outputs; coordinate decoding and NMS remain outside the exported graph.
  • Float16 and float32 can retain different people near the score threshold.

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).