Models · Dense matching

XFeat

XFeat (Verlab, CVPR 2024): fast sparse keypoints and descriptors, matched by mutual nearest neighbours or LighterGlue; output as keypoints plus index pairs (the form COLMAP takes). Apache-2.0 code and weights.

experimental Fork of verlab/accelerated_features@e92685f · backends: portable, kernels

Checkpoint

Paper 2024 Added in 0.2.0 Evaluated with 0.2.0

Portable, RTX 5060 Ti 16 GB 7.17 ms 768x1024, fp32
vs the faster original 1.2× original fp16: 8.89 ms
Peak memory 122 MiB original: 128 MiB
Accuracy vs original fp32 ±0.0 portable AUC@5°, MegaDepth-1500 (29.1 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("xfeat")  # 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 xfeat

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 fp32. Portable runs fp32 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 fp16torch.autocast(float16) around the call; upstream XFeat.detectAndCompute + mnn, top_k 4096fp16 8.89–128 MiB/174 MiB0.30.32
Portable fp32fp32not measured ––––
Portablefp32 7.171.2×122 MiB/154 MiB0.20.74
Kernelsnot measured ––––
Portable + torch.compile 5.651.6×186 MiB/212 MiB31 (build)–
TensorRTplanned for 0.2.1fails––––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-12 to 2026-11-13. 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, 2026-11-13Originalfp3245.2 1,549 MiB process–768x1024
Portablefp3241.7 1,471 MiB processmatches768x1024
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
Apache-2.0 LICENSE
Weights
Apache-2.0 source
Training data

Our reading of the licences, not legal advice.

Weights

xfeat.pt 6.2 MB· hosted on GitHub

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

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 fp16not measured
Portable fp32fp3226.242.257.5 1500 15.5
Portablefp3226.242.257.5 1500 15.5
Kernelsnot measured

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. Measured on RTX 5060 Ti, torch 2.14.1+cu130, fastforks ec7b1307+por, 2026-11-12.

MegaDepth-1500

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

RundtypeAUC@5° ↑AUC@10° ↑AUC@20° ↑Δ AUC@5° vs fp32 originalpairsms / pair
Original fp32fp3229.146.160.8 reference 300 15.4
Original fp16fp1628.444.660.1 −0.8 300 12.6
Portable fp32fp3229.146.160.8 ±0.0 300 15.7
Portablefp3229.146.160.8 ±0.0 300 15.7
Kernelsnot 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. The same 300 pairs for all modes, so float32 and bfloat16 compare pair for pair. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks ec7b1307+por, 2026-11-12.

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| <= 3e-05 vs the original. GPU, float32 (the default), portable and kernels: mean |diff| <= 0.0001 and p99 <= 0.001 (absolute error of the output keypoints1).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.00003 not recorded
cpu_min_recall≤ 0.99 not recorded
cpu_min_match≤ 0.99 not recorded
GPU default dtypes, mean |diff|≤ 0.0001 0
GPU default dtypes, p99 |diff|≤ 0.001 0
gpu_min_recall≤ 0.999 not recorded
gpu_min_match≤ 0.993 not recorded

Recorded for xfeat on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: keypoint_nms

Details

XFeat (Potje et al., CVPR 2024): lightweight sparse feature matching. A small CNN finds up to 4096 keypoints per image with 64-dimensional descriptors; they are matched as mutual nearest neighbours or with LighterGlue, a smaller LightGlue trained for XFeat. Returns both keypoint sets, the index pairs of the matches and their confidence.

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 (1,047 MiB allocated)
Options
  • max_keypoints=4096>= 1· changes the output
  • long_side=0>= 0· changes the output
Outputs
keypoints1, keypoints2, pairs, scores
Checkpoint
matcher="mnn", detection_threshold=0.05
CUDA kernels
  • keypoint_nms
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 keypoint counts differ per image.
  • No exportable module: the keypoint counts are data-dependent (`torch.compile` runs with graph breaks; ONNX / TensorRT export fails).
  • `xfeat` (mutual nearest neighbours) has no learned confidence: its scores are the descriptors' cosine similarities.
Full reference, generated from the repository

XFeat (Verlab, CVPR 2024): fast sparse keypoints and descriptors, matched by mutual nearest neighbours or LighterGlue; output as keypoints plus index pairs (the form COLMAP takes). Apache-2.0 code and weights.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
xfeatsynthetic (768x1024)original
portable
kernels
portable + torch.compile
TensorRT fp16
8.89
7.17
–
5.65
planned for 0.2.1
128 / 174
122 / 154
–
186 / 212
–
xfeat-lightergluesynthetic (768x1024)original
portable
kernels
portable + torch.compile
TensorRT fp16
23.6
21.6
–
16.2
planned for 0.2.1
448 / 478
445 / 476
–
262 / 278
–
Details
  • Original code: https://github.com/verlab/accelerated_features at e92685f57f83
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
  • Licence evidence: https://github.com/verlab/accelerated_features/blob/e92685f57f8318b18725c5c8c0bd28c7fe188d9a/LICENSE ; https://github.com/verlab/accelerated_features/blob/e92685f57f8318b18725c5c8c0bd28c7fe188d9a/README.md#license
  • Training data: MegaDepth (Internet photo collections with SfM / MVS depth), COCO 2017 (train, instance annotations) (XFeat: MegaDepth and synthetic warps of a 20k-image COCO 2017 subset (README training section). LighterGlue: trained with GlueFactory (README; MegaDepth, data not otherwise stated).)
  • Options: max_keypoints = 4096 (>= 1), long_side = 0 (>= 0)
  • Kernel libraries: keypoint_nms
  • Variants: xfeat, xfeat-lighterglue
  • Not in this release (planned for 0.2.1): xfeat ONNX / TensorRT export: torch.export stops at a data-dependent branch (NMS / keypoint selection); xfeat-lighterglue ONNX / TensorRT export: torch.export stops at a data-dependent branch (NMS / keypoint selection)
  • 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): xfeat 3000x4000 (1.0 GiB peak allocated, 1.2 reserved); xfeat-lighterglue 3000x4000 (1.0 GiB peak allocated, 1.2 reserved)
  • xfeat 768x1024, first call: portable + torch.compile 31 s (cold Inductor cache).
  • xfeat-lighterglue 768x1024, first call: portable + torch.compile 40 s (cold Inductor cache).
  • xfeat original 768x1024: torch.autocast(float16) around the call; upstream XFeat.detectAndCompute + mnn, top_k 4096
  • xfeat-lighterglue original 768x1024: torch.autocast(float16) around the call; upstream XFeat.detectAndCompute + lighterglue, top_k 4096
  • xfeat load / first call: original 1.51 s / 0.629 s, portable 0.16 s / 0.74 s
  • xfeat-lighterglue load / first call: original 1.46 s / 0.692 s, portable 0.25 s / 0.681 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. The same 300 pairs for all modes, so float32 and bfloat16 compare pair for pair. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portablenote
xfeatMegaDepth-1500, relative pose AUC@5 / @10 / @20 degrees26.2 / 42.2 / 57.5
xfeatMegaDepth-1500, every 5th pair, relative pose AUC@5 / @10 / @20 degrees29.1 / 46.1 / 60.828.4 / 44.6 / 60.129.1 / 46.1 / 60.8
xfeat-lighterglueMegaDepth-1500, relative pose AUC@5 / @10 / @20 degrees39.7 / 57.5 / 72.8
xfeat-lighterglueMegaDepth-1500, every 5th pair, relative pose AUC@5 / @10 / @20 degrees42.0 / 59.4 / 74.342.7 / 60.1 / 74.142.0 / 59.4 / 74.3

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks ec7b1307+port/p1-match, 2026-11-12.

Parity

CPU, float32: max |diff| <= 3e-05 vs the original. GPU, float32 (the default), portable and kernels: mean |diff| <= 0.0001 and p99 <= 0.001 (absolute error of the output keypoints1).

variantkernels error vs original, mean / p99GPU, torchkernels used
xfeat0.0 / 0.0RTX 5060 Ti, torch 2.10.0+cu130keypoint_nms@3cfcd62bc9a5:sm_120
xfeat-lighterglue0.0 / 0.0RTX 5060 Ti, torch 2.10.0+cu130keypoint_nms@3cfcd62bc9a5: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 keypoint counts differ per image.
  • No exportable module: the keypoint counts are data-dependent (torch.compile runs with graph breaks; ONNX / TensorRT export fails).
  • xfeat (mutual nearest neighbours) has no learned confidence: its scores are the descriptors' cosine similarities.

Example output

No example rendered for this checkpoint yet.