Research map / Architectures and efficiency / CNN design
XFeat: Accelerated Features for Lightweight Image Matching
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces XFeat, a lightweight CNN for keypoint detection, local description and sparse/semi-dense matching that runs in real time on CPU, trained with distillation of ALIKE keypoints.
Problem
Accurate local feature extraction and image matching are computationally expensive, and existing deep local features are too slow for resource-constrained devices.
Approach
A featherweight CNN that keeps early-layer resolution high with few channels (tripling channels as resolution halves), a separate 8x8-grid keypoint branch distilled from ALIKE-Tiny, and a lightweight MLP match refinement module on coarse 1/8-resolution descriptors enabling semi-dense matching.
Key results
- On MegaDepth-1500, XFeat runs at 27.1 FPS on a laptop CPU versus 5.3 for ALIKE (about 5x faster), with AUC@20 of 67.7 vs 71.4 for ALIKE.
- XFeat* (semi-dense, 10k keypoints) reaches AUC@20 of 77.1 and Acc@10 of 85.1 on MegaDepth-1500, outperforming DISK* (75.3 and 81.3) at roughly 16x speed (19.2 vs 1.2 FPS).
- On Aachen day-night localization, XFeat is comparable to SuperPoint and DISK while being at least 9x faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use semi-dense matching rather than sparse matching | XFeat · Megadepth-1500 | AUC@20°: 67.7 -> 77.1 | Helped |
| Use semi-dense matching rather than sparse matching | XFeat · ScanNet-1500 | AUC@20°: 47.8 -> 50.3 | Helped |
| Remove synthetic warps from training (ablation) | XFeat · Megadepth-1500 | AUC@5°: 42.6 -> 41.5 | Hurt |
| Remove synthetic warps from training (ablation) | XFeat* · Megadepth-1500 | AUC@5°: 50.2 -> 33.9 | Hurt |
| Halve channels in the last three convolutional blocks to 32 (ablation) | XFeat · Megadepth-1500 | AUC@5°: 42.6 -> 37.4 | Hurt |
| Halve channels in the last three convolutional blocks to 32 (ablation) | XFeat* · Megadepth-1500 | AUC@5°: 50.2 -> 40.7 | Hurt |
| Replace the parallel keypoint head with joint keypoint extraction from descriptor embeddings (ablation) | XFeat · Megadepth-1500 | AUC@5°: 42.6 -> 42.9 | No clear effect |
| Replace the parallel keypoint head with joint keypoint extraction from descriptor embeddings (ablation) | XFeat* · Megadepth-1500 | AUC@5°: 50.2 -> 39.7 | Hurt |
| Remove match refinement (ablation) | XFeat* · Megadepth-1500 | AUC@5°: 50.2 -> 38.6 | Hurt |
Surprising findings
- Naively pruning channels or halving final channel widths severely degrades accuracy, so the gains come from redistributing depth rather than uniform shrinking.
- Jointly training the keypoint and descriptor heads in a compact backbone degraded semi-dense matching performance.
Limitations
- Semi-dense matching (XFeat*) runs at roughly 19 FPS and uses multi-scale inference, costing more than sparse mode.
- Sparse accuracy is below ALIKE on some MegaDepth metrics at tighter thresholds.
- Keypoint branch depends on ALIKE-Tiny teacher labels rather than a self-supervised objective.
Experimental setup
Datasets: MegaDepth COCO (synthetically warped) ALIKE keypoint teacher labels
Benchmarks: MegaDepth-1500 ScanNet-1500 HPatches Aachen day-night
Metrics: AUC@5/10/20 Acc@10 MIR #inliers FPS MHA localization accuracy at 0.25m/0.5m/5m and 2/5/10 deg
Baselines: SuperPoint DISK SiLK ORB ZippyPoint ALIKE LightGlue
Builds on: SuperPoint ALIKE LoFTR DISK
Closest papers by meaning
- ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses (NeurIPS 2024 · similarity 0.85)
- xT: Nested Tokenization for Larger Context in Large Images (ICML 2024 · similarity 0.85)
- JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba (CVPR 2025 · similarity 0.85)
- UFM: A Simple Path towards Unified Dense Correspondence with Flow (NeurIPS 2025 · similarity 0.85)
- GIM: Learning Generalizable Image Matcher From Internet Videos (ICLR 2024 · similarity 0.84)
- Sample- and Parameter-Efficient Auto-Regressive Image Models (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.