Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation
Authors: Xiaolong Wang, Lei Yu, Yingying Zhang, Jiangwei Lao, Lixiang Ru, Liheng Zhong, Jingdong Chen, Yu Zhang, Ming Yang
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method achieves higher accuracy compared to previous semi-dense matchers. Meanwhile, our dense matching results exhibit similar end-point-error accuracy compared to previous dense matchers while maintaining semi-dense efficiency. ... We conduct comprehensive experiments on the Lo FTR and ASpan Former models, demonstrating that our method significantly enhances model performance, even reaching state-of-the-art levels for semi-dense matching methods. ... The experimental results indicate that our method significantly outperforms other semi-dense approaches and achieves similar results to dense methods. |
| Researcher Affiliation | Collaboration | Xiaolong Wang1* , Lei Yu2*, Yingying Zhang2, Jiangwei Lao2, Lixiang Ru2, Liheng Zhong2, Jingdong Chen2, Yu Zhang1 , Ming Yang2 1College of Control Science and Engineering, Zhejiang University 2 Ant Group EMAIL, EMAIL |
| Pseudocode | No | The paper describes the method using figures (e.g., Figure 2) and prose within the 'Method' section, but it does not include any explicitly labeled pseudocode block or algorithm. |
| Open Source Code | No | The paper mentions using 'Open CV s RANSAC implementation' and refers to 'using the official code' for other methods like Ro Ma, but does not provide a specific link or explicit statement about releasing the source code for their own proposed method, Homo Matcher. |
| Open Datasets | Yes | We employ the Mega Depth dataset (Li and Snavely 2018) and Scan Net (Dai et al. 2017) datasets to validate our model s matching capability for relative pose estimation in both outdoor and indoor settings. ... The HPatches dataset (Balntas et al. 2017) provides multiple sets of sequential images of the same scenes under varying viewpoints and illumination conditions |
| Dataset Splits | Yes | We follow the training and validation splits used in previous methods (Sun et al. 2021), with the validation set comprising 1500 randomly selected image pairs from scenes 0015 and 0022. Images are resized with the longest edge to 832 and 1152 for training and validation, respectively. Scan Net is a large-scale indoor dataset ... We evaluate using the same 1500 test image pairs as (Sarlin et al. 2020), with all test images resized to 480 640. |
| Hardware Specification | Yes | The model training utilizes the Adam optimizer with a learning rate of 1 10 3 and a batch size of 8 for 30 epochs on 8 V-100 GPUs. ... On a single V100 GPU with Mega Depth images resized to 1152 resolution, ASpan Homo runs in 442 ms, the heavy version in 697 ms, while Ro Ma, using the official code, takes 1527 ms. |
| Software Dependencies | No | This computation is performed using Open CV s RANSAC implementation with settings consistent with (Sun et al. 2021). The model training utilizes the Adam optimizer... While specific software like OpenCV and Adam optimizer are mentioned, no version numbers are provided for these or any other libraries. |
| Experiment Setup | Yes | The threshold for obtaining coarse matches is set at θc = 0.2. The size of patches cropped from the fine-level features is w = 9. The correlation search radius during iterative homography estimation is r = 1, with K = 3 iterations. The densification radius used during loss calculation is re = 2. The model training utilizes the Adam optimizer with a learning rate of 1 10 3 and a batch size of 8 for 30 epochs on 8 V-100 GPUs. |