CasP 🪜

Improving Semi-Dense Feature Matching Pipeline Leveraging
Cascaded Correspondence Priors for Guidance

ICCV 2025

Peiqi Chen^1* · Lei Yu^2* · Yi Wan^1† Yingying Pei¹ · Xinyi Liu¹ · Yongxiang Yao¹
Yingying Zhang² · Lixiang Ru² · Liheng Zhong² · Jingdong Chen² · Ming Yang² · Yongjun Zhang^1†

¹Wuhan University ²Ant Group
*Equal contribution †Corresponding author

CasP decomposes the matching stage into two progressive layers, with the former layer providing the one-to-many priors that constrain the search range of the latter.

## Update - **[2025-08]** The inference code has been released, and an online demo is available on [Hugging Face Spaces](https://huggingface.co/spaces/pq-chen/CasP), which includes the outdoor model trained on MegaDepth and the fine-tuned model on MINIMA. - **[2025-07]** CasP has been accepted to ICCV 2025 as a **highlight** paper. > **Note:** Due to disclosure restrictions imposed by the funding agency, model weights are only available through the demo. For any non-commercial request, please contact the corresponding author at yi.wan@whu.edu.cn. ## Introduction This repository hosts the official implementation of CasP, a cascaded semi-dense feature matching pipeline designed for superior accuracy and efficiency. ## Benchmark

[MegaDepth-Synthetic-1500 (MD-Syn-1500)]

Here are the results as Area Under the Curve (AUC) of the relative pose error at 10 degree:

| Method | Fine-tuned | RGB-Infrared | RGB-Depth | RGB-Normal | RGB-Event | RGB-Sketch | RGB-Paint | |----------------|------------|--------------|-----------|------------|-----------|------------|-----------| | LoFTR | ❌ | 12.58 | 0.44 | 12.07 | 12.43 | 54.82 | 12.22 | | ELoFTR | ❌ | 14.59 | 0.79 | 21.67 | 20.39 | 61.09 | 25.11 | | CasP | ❌ | **22.53** | **1.20** | **30.25** | **35.51** | **62.92** | **39.70** | | MINIMA_LoFTR | ✅ | 32.36 | 28.81 | 44.26 | 32.74 | 53.54 | 15.45 | | MINIMA_ELoFTR | ✅ | 26.36 | 32.26 | 47.47 | 30.72 | 59.63 | 27.02 | | MINIMA_CasP | ✅ | **43.87** | **40.55** | **53.64** | **40.06** | **60.30** | **40.76** |

[Zero-Shot Cross-Modality Matching]

Here are visualizations of image registration across unseen modalities, produced by our model fine-tuned on MINIMA:

Optical-Point Cloud

Optical-SAR

Optical-Vector Map

## Citation If you find any of the ideas or the code useful for your research, please consider citing our paper: ```bibtex @inproceedings{chen2025casp, title={CasP: Improving Semi-Dense Feature Matching Pipeline Leveraging Cascaded Correspondence Priors for Guidance}, author={Chen, Peiqi and Yu, Lei and Wan, Yi and Pei, Yingying and Liu, Xinyi and Yao, Yongxiang and Zhang, Yingying and Ru, Lixiang and Zhong, Liheng and Chen, Jingdong and Yang, Ming and Zhang, Yongjun}, booktitle={ICCV}, year={2025} } ``` ## License The pre-trained models of CasP and the code provided in this repository are released under the [Apache-2.0 license](LICENSE).