Models · Dense matching

DeDoDe v2

DeDoDe v2 (Chalmers, Linköping University; 3DV 2024): sparse keypoints from a detector network, descriptions from a separate one (VGG19, plus DINOv2 for -G), dual-softmax matches as keypoint index pairs (the form COLMAP takes). MIT; DINOv2 Apache-2.0.

experimental Fork of Parskatt/DeDoDe@6d15618 · backends: portable, kernels

Checkpoint

Paper 2024 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 189 ms 768x1024, fp16
vs the faster original 1.7× original fp16: 318 ms
Peak memory 1,197 MiB original: 1,734 MiB
Accuracy vs original fp32 – original not run on the test set

Use it

pip install fastforks

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

from fastforks import matching

model = matching.load("dedode-v2-b")  # kernels on a supported GPU, else portable PyTorch
result = model(image_a, image_b)
result.keypoints1  # also: keypoints2, pairs, scores
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 dedode-v2-b

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 DeDoDe-L v2 + descriptor B, 10000 keypointsfp16 318–1,734 MiB/1,992 MiB2.61.54
Portable fp32fp32not measured ––––
Portablefp16 2461.3×1,535 MiB/1,654 MiB0.70.75
Kernelsfp16 1891.7×1,197 MiB/1,246 MiB0.60.84
Portable + torch.compile 1641.9×1,893 MiB/2,434 MiB50 (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, 2026-11-17Originalfp321711 8,024 MiB process–768x1024
Portablefp321464 8,011 MiB processmatches768x1024
Kernelsfp16not implemented–768x1024
Kernelsfp32not 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 LICENSELICENSE-DINOv2
Weights
MIT source
Training data

Our reading of the licences, not legal advice.

Weights

dedode_detector_L_v2.pth 58.5 MB· hosted on GitHub
dedode_descriptor_B.pth 54.3 MB· hosted on GitHub

fastforks downloads and checks these files on first use (fastforks download dedode-v2-b).

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

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

RundtypeAUC@5° ↑AUC@10° ↑AUC@20° ↑pairsms / pair
Original fp32not measured
Original fp16fp1657.371.582.5 150 318
Portable fp32fp3256.072.082.7 150 465
Portablefp1654.671.883.1 150 246
Kernelsfp1657.874.384.2 150 189

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 fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks d073b335, 2026-11-18; portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks aee4b565, 2026-11-13; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16.

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| <= 0.011 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.002 and p99 <= 0.001 (absolute error of the output keypoints1).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.011 not recorded
cpu_min_recall≤ 0.99 not recorded
cpu_min_match≤ 0.99 not recorded
GPU default dtypes, mean |diff|≤ 0.002 0.002
GPU default dtypes, p99 |diff|≤ 0.001 0
gpu_min_recall≤ 0.9986 not recorded
gpu_min_match≤ 0.986 not recorded
gpu_mean_kernels≤ 0.003 not recorded
gpu_min_match_kernels≤ 0.98 not recorded

Recorded for dedode-v2-b on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-15, with these kernels: cudnn_hacccutlass_convigev

Details

DeDoDe v2 (Edstedt et al., 3DV 2024; detector v2: CVPR Workshops 2024): detect, don't describe, describe, don't detect. A VGG19 network with convolutional refiners finds 10000 keypoints per image (the detector, retrained in v2), a separate descriptor network (VGG19, or VGG19 + a frozen DINOv2 ViT-L/14 for -G) describes them, and the descriptions are matched as mutual maxima of a dual softmax. Returns both keypoint sets, the index pairs of the matches and their probability.

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 (1,983 MiB allocated)
Options
  • size=784>= 56· changes the output
  • max_keypoints=10000>= 1· changes the output
Outputs
keypoints1, keypoints2, pairs, scores
Checkpoint
descriptor="B"
CUDA kernels
  • igevIGEV stereo: convolution epilogues, correlation lookups, GRU gates
  • cutlass_convCUTLASS convolutions with the recurrent update (GRU) fused in
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
Apple (MPS)
fp32

Limitations

  • Sparse output: no dense warp. `result.matches()` gives the matched points for `estimate_pose`; `pairs` are the index pairs pycolmap / COLMAP databases take.
  • One image pair per call internally (batches run pair by pair); the match counts differ per pair.
  • No exportable module: the match counts are data-dependent (`torch.compile` runs with graph breaks; ONNX / TensorRT export fails).
  • Images are resized to a `size` x `size` square (default 784, as upstream) whatever their aspect ratio; the dual softmax holds a `max_keypoints`^2 matrix (10000 keypoints: 400 MB per float32 copy).
  • Upright only: the released descriptors are not trained for in-plane rotations.
Full reference, generated from the repository

DeDoDe v2 (Chalmers, Linköping University; 3DV 2024): sparse keypoints from a detector network, descriptions from a separate one (VGG19, plus DINOv2 for -G), dual-softmax matches as keypoint index pairs (the form COLMAP takes). MIT; DINOv2 Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
dedode-v2-bsynthetic (768x1024)original
portable
kernels
portable + torch.compile
TensorRT fp16
318
246
189
164
planned for 0.2.1
1734 / 1992
1535 / 1654
1197 / 1246
1893 / 2434
–
dedode-v2-gsynthetic (768x1024)original
portable
kernels
portable + torch.compile
TensorRT fp16
666
592
332
330
planned for 0.2.1
2335 / 2472
2142 / 2334
1787 / 1984
2881 / 3316
–
Details
  • Original code: https://github.com/Parskatt/DeDoDe at 6d156183f4dc
  • Code licence: MIT (permissive); files: LICENSE, LICENSE-DINOv2
  • Weights licence: MIT
  • Licence evidence: https://github.com/Parskatt/DeDoDe/blob/6d156183f4dc84cd704ae779eebc8350995c5b06/README.md#license ; https://github.com/Parskatt/DeDoDe/blob/6d156183f4dc84cd704ae779eebc8350995c5b06/LICENSE
  • Training data: MegaDepth (Internet photo collections with SfM / MVS depth) (Detector v2 and descriptors: MegaDepth (README training section, papers); DINOv2 (-G, frozen): LVD-142M (Meta).)
  • Options: size = 784 (>= 56), max_keypoints = 10000 (>= 1)
  • Kernel libraries: igev, cutlass_conv, cudnn_hacc
  • Variants: dedode-v2-b, dedode-v2-g
  • Not in this release (planned for 0.2.1): dedode-v2-b TensorRT export: TensorRT TopK supports K <= 3840; the keypoint top-k is larger; dedode-v2-g TensorRT export: TensorRT TopK supports K <= 3840; the keypoint top-k is larger
  • Outputs: keypoints1, keypoints2, pairs, scores
  • Input size: any H x W >= 32, padded to a multiple of 32
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): dedode-v2-b 3000x4000 (1.9 GiB peak allocated, 3.8 reserved); dedode-v2-g 3000x4000 (2.5 GiB peak allocated, 4.5 reserved)
  • dedode-v2-b 768x1024, first call: portable + torch.compile 50 s (cold Inductor cache).
  • dedode-v2-g 768x1024, first call: portable + torch.compile 86 s (cold Inductor cache).
  • dedode-v2-b original 768x1024: the authors' mixed-precision option (float16); upstream DeDoDe-L v2 + descriptor B, 10000 keypoints
  • dedode-v2-g original 768x1024: the authors' mixed-precision option (float16); upstream DeDoDe-L v2 + descriptor G, 10000 keypoints
  • dedode-v2-b load / first call: original 3.73 s / 1.099 s, portable 0.67 s / 0.753 s, kernels 0.55 s / 0.84 s
  • dedode-v2-g load / first call: original 6.36 s / 1.509 s, portable 1.12 s / 1.038 s, kernels 3.01 s / 0.675 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 fp16portableportable fp32kernelsnote
dedode-v2-bMegaDepth-1500, every 10th pair, relative pose AUC@5 / @10 / @20 degrees57.3 / 71.5 / 82.554.6 / 71.8 / 83.156.0 / 72.0 / 82.757.8 / 74.3 / 84.2

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks aee4b565, 2026-11-13; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks d073b335, 2026-11-18; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16.

Parity

CPU, float32: max |diff| <= 0.011 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.002 and p99 <= 0.001 (absolute error of the output keypoints1).

variantkernels error vs original, mean / p99GPU, torchkernels used
dedode-v2-b0.002 / 0.0RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
dedode-v2-g0.002 / 0.0RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120

Limitations

  • Sparse output: no dense warp. result.matches() gives the matched points for estimate_pose; pairs are the index pairs pycolmap / COLMAP databases take.
  • One image pair per call internally (batches run pair by pair); the match counts differ per pair.
  • No exportable module: the match counts are data-dependent (torch.compile runs with graph breaks; ONNX / TensorRT export fails).
  • Images are resized to a size x size square (default 784, as upstream) whatever their aspect ratio; the dual softmax holds a max_keypoints^2 matrix (10000 keypoints: 400 MB per float32 copy).
  • Upright only: the released descriptors are not trained for in-plane rotations.

Example output

No example rendered for this checkpoint yet.