Models · Dense matching

RoMa

RoMa (2024): dense matching between two photos of one scene from any viewpoints: where each pixel of the first image is in the second, and how sure the model is that it is visible there. Built on Meta's DINOv2 and a VGG19. MIT code and weights; the DINOv2 weights are Apache-2.0.

experimental Fork of Parskatt/RoMa@77f8d68 · backends: portable, kernels

Checkpoint romav1

Paper 2023 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 356 ms 768x1024, fp16
vs the faster original 1.1× original fp16: 386 ms
Peak memory 1,720 MiB original: 3,887 MiB
Accuracy vs original fp32 −0.1 kernels AUC@5°, MegaDepth-1500 (67.6 original)

Use it

pip install fastforks

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

from fastforks import matching

model = matching.load("romav1")  # kernels on a supported GPU, else portable PyTorch
result = model(image_a, image_b)
result.warp  # also: certainty
result.backend  # what ran: mode, device, dtype, kernels

points1, points2 = result.matches(5000)  # matched pixels, sampled where the model is certain
pose = matching.estimate_pose(points1, points2, K1, K2)  # relative pose on the GPU (RANSAC); K = intrinsics [3, 3]
pose.R, pose.t, pose.inliers  # rotation, unit translation, inlier mask
fastforks bench romav1

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, 768x1024 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32not measured ––––
Original fp16the authors' mixed-precision option (float16); upstream RoMa v1 match, one direction; local correlation: upstream's plain-torch path (use_custom_corr off)fp16 386–3,887 MiB/4,094 MiB4.90.47
Portable fp32fp32not measured ––––
Portablefp16 3771.0×1,720 MiB/1,840 MiB2.30.92
Kernelsfp16 3561.1×1,720 MiB/1,860 MiB3.21.03
Portable + torch.compile 2471.6×1,348 MiB/1,476 MiB698 (build)–
TensorRTplanned for 0.2.1fails––––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-13 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. 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, 2026-11-13Originalfp3212208 12.2 GiB process–768x1024
Portablefp323050 7,979 MiB processmatches768x1024
Kernelsfp16not implemented–768x1024

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
MIT LICENSE
Weights
MIT AND Apache-2.0 source
Training data
  • The DINOv2 ViT-L/14 weights are the second weight file (Meta's release, Apache-2.0).

Our reading of the licences, not legal advice.

Weights

roma_outdoor.pth 446 MB· hosted on GitHub
dinov2_vitl14_pretrain.pth 1.2 GB· hosted on Meta (fbaipublicfiles)

fastforks downloads and checks these files on first use (fastforks download romav1).

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.

MegaDepth-1500

relative pose AUC@5 / @10 / @20 degrees

RundtypeAUC@5° ↑AUC@10° ↑AUC@20° ↑pairsms / pair
Original fp32not measured
Original fp16fp1662.676.686.3 1500 1477
Portable fp32not measured
Portablefp1662.676.686.4 1500 1408
Kernelsfp1662.476.486.2 1500 1343

The original code was not run on this dataset, so the table cannot show the distance to it.

Dense matchers: 5000 matches per pair sampled from the warp by certainty (upstream's balanced KDE sampling); sparse matchers: all their matches of one call (LightGlue, XFeat: LightGlue's MegaDepth-1500 setting; LoFTR, DeDoDe: their card defaults). OpenCV essential-matrix RANSAC (0.5 px threshold, normalised by the mean focal length); AUC of the larger of the rotation and translation-direction errors. Original = the authors’ code, same checkpoint and protocol. Measured on original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks f218aa9d, 2026-11-13.

MegaDepth-1500

every 10th pair, relative pose AUC@5 / @10 / @20 degrees

RundtypeAUC@5° ↑AUC@10° ↑AUC@20° ↑Δ AUC@5° vs fp32 originalpairsms / pair
Original fp32fp3267.680.188.2 reference 150 2115
Original fp16fp1666.479.588.1 −1.2 150 1477
Portable fp32fp3266.579.788.0 −1.1 150 2812
Portablefp1668.280.888.8 +0.6 150 1408
Kernelsfp1667.580.088.2 −0.1 150 1343

Dense matchers: 5000 matches per pair sampled from the warp by certainty (upstream's balanced KDE sampling); sparse matchers: all their matches of one call (LightGlue, XFeat: LightGlue's MegaDepth-1500 setting; LoFTR, DeDoDe: their card defaults). OpenCV essential-matrix RANSAC (0.5 px threshold, normalised by the mean focal length); AUC of the larger of the rotation and translation-direction errors. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13; portable fp32, portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks f218aa9d, 2026-11-13.

Relative pose on the GPU

A dense matcher is usually followed by RANSAC on its matches to get the camera motion. fastforks does that step on the GPU too: result.matches(n) samples matched points where the model is certain, and matching.estimate_pose runs essential-matrix RANSAC on them, many image pairs per call, with no copy to the CPU. Like the models it has a portable (plain PyTorch) and a kernels (CUDA) backend.

SolverRuns onThreshold (px)Pose AUC@5 / @10 / @20Pairs per second
fastforks estimate_pose, kernelsGPU, batches of 500 pairs1.069.9 / 81.5 / 89.318,446
fastforks estimate_pose, portableGPU, batches of 500 pairs1.069.6 / 81.2 / 89.1418
PoseLib 2.0.5CPU, 16 processes1.070.0 / 81.6 / 89.5325
OpenCV USAC (MAGSAC++)CPU, 16 processes0.565.7 / 78.6 / 87.3271
OpenCV RANSACCPU, 16 processes0.562.9 / 76.9 / 86.4323

MegaDepth-1500, RoMa v2 matches (5000 per pair), 5 samplings, the same matches for every solver, on one machine: RTX 5060 Ti and a Ryzen 7 5800X (16 threads). Each solver runs at the inlier threshold where it is most accurate. Pairs per second is wall-clock throughput over all 7500 poses: the CPU solvers run one pair per call in 16 single-threaded processes, the GPU solver takes 500 pairs per call. One pair alone takes about 1.7 ms with the kernels, 29 ms with PoseLib and 24 ms with OpenCV RANSAC on one core. The accuracy table above stays on OpenCV RANSAC, as in the papers, so it is comparable with published numbers; this table is the only place the GPU solver is used.

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| <= 4.3e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.00079 and p99 <= 0.0042 (absolute error of the output warp).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000043 not recorded
gpu_min_score≤ 0.5 not recorded
GPU default dtypes, mean |diff|≤ 0.00079 0.00052
GPU default dtypes, p99 |diff|≤ 0.0042 0.00269

Recorded for romav1 on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: fastbb

Details

RoMa (Edstedt, Sun, Bökman, Wadenbäck, Felsberg, CVPR 2024): dense feature matching between two images of any viewpoint. A frozen DINOv2 ViT-L/14 and a VGG19-BN describe both images; a Gaussian process over the coarse DINOv2 features and a transformer classifier give a coarse warp, and convolutional refiners (local correlation at strides 16, 8, 4) sharpen it at strides 16 to 1, at a coarse square size and again at a fine size. Returns, for every pixel of the first image, its position in the second (pixels) and the probability that it is visible there.

Input / output
Input: two RGB images of one scene. Output: for every pixel of the first image its position in the second (x, y in pixels) and a certainty that the pixel is visible in both.
What the numbers mean, and which models need the camera focal length

Sparse matchers (LightGlue) return the keypoints of both images and the index pairs of the matched ones instead. To reconstruct camera poses from an image folder with these matchers, see [structure from motion](../sfm.md).

Input size
any H x W >= 32, padded to a multiple of 32; tested up to 3000x4000 (1,989 MiB allocated)
Options
  • coarse_size=560>= 14· changes the output
  • fine_size=864>= 0· changes the output
Outputs
warp, certainty
CUDA kernels
  • fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbones
Apple (MPS)
fp32

Limitations

  • One direction (the first image's half of upstream's symmetric call); upstream's sampling / keypoint helpers are not included; tools/eval_matching.py has the MegaDepth-1500 sampling.
  • Both images are resized to squares (upstream's behaviour), whatever their aspect ratio; upstream resizes PIL images, this runner float tensors (bicubic, antialiased).
Full reference, generated from the repository

RoMa (2024): dense matching between two photos of one scene from any viewpoints: where each pixel of the first image is in the second, and how sure the model is that it is visible there. Built on Meta's DINOv2 and a VGG19. MIT code and weights; the DINOv2 weights are Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
romav1synthetic (768x1024)original
portable
kernels
portable + torch.compile
TensorRT fp16
386
377
356
247
planned for 0.2.1
3887 / 4094
1720 / 1840
1720 / 1860
1348 / 1476
–
Details
  • Original code: https://github.com/Parskatt/RoMa at 77f8d6880352
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT AND Apache-2.0
  • Licence evidence: https://github.com/Parskatt/RoMa/blob/77f8d68803526dcddfd9b7a46bc76125bdc25f15/README.md#license
  • Training data: MegaDepth (Internet photo collections with SfM / MVS depth) (roma_outdoor: MegaDepth (upstream README / paper). The frozen DINOv2 was pretrained on LVD-142M (not released).)
  • Licence note: The DINOv2 ViT-L/14 weights are the second weight file (Meta's release, Apache-2.0).
  • Options: coarse_size = 560 (>= 14), fine_size = 864 (>= 0)
  • Kernel libraries: fastbb
  • Variants: romav1
  • Not in this release (planned for 0.2.1): romav1 ONNX / TensorRT export: the ONNX converter fails on the exported program (ConversionError, step 3/3)
  • Outputs: warp, certainty
  • Input size: any H x W >= 32, padded to a multiple of 32
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): romav1 3000x4000 (1.9 GiB peak allocated, 2.2 reserved)
  • romav1 768x1024, first call: portable + torch.compile 698 s (cold Inductor cache).
  • romav1 original 768x1024: the authors' mixed-precision option (float16); upstream RoMa v1 match, one direction; local correlation: upstream's plain-torch path (use_custom_corr off)
  • romav1 load / first call: original 6.28 s / 2.57 s, portable 2.34 s / 0.918 s, kernels 3.17 s / 1.026 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Dense matchers: 5000 matches per pair sampled from the warp by certainty (upstream's balanced KDE sampling); sparse matchers: all their matches of one call (LightGlue, XFeat: LightGlue's MegaDepth-1500 setting; LoFTR, DeDoDe: their card defaults). OpenCV essential-matrix RANSAC (0.5 px threshold, normalised by the mean focal length); AUC of the larger of the rotation and translation-direction errors. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
romav1MegaDepth-1500, relative pose AUC@5 / @10 / @20 degrees62.6 / 76.6 / 86.362.6 / 76.6 / 86.462.4 / 76.4 / 86.2
romav1MegaDepth-1500, every 10th pair, relative pose AUC@5 / @10 / @20 degrees67.6 / 80.1 / 88.266.4 / 79.5 / 88.168.2 / 80.8 / 88.866.5 / 79.7 / 88.067.5 / 80.0 / 88.2

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks f218aa9d, 2026-11-13.

Parity

CPU, float32: max |diff| <= 4.3e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.00079 and p99 <= 0.0042 (absolute error of the output warp).

variantkernels error vs original, mean / p99GPU, torchkernels used
romav10.00052 / 0.00269RTX 5060 Ti, torch 2.14.1+cu130fastbb@7d058322634d:sm_120

Limitations

  • One direction (the first image's half of upstream's symmetric call); upstream's sampling / keypoint helpers are not included; tools/eval_matching.py has the MegaDepth-1500 sampling.
  • Both images are resized to squares (upstream's behaviour), whatever their aspect ratio; upstream resizes PIL images, this runner float tensors (bicubic, antialiased).

Example output

No example rendered for this checkpoint yet.