Method Towards CVPR 2021 SimLocMatch Challenge

Xiaopeng Bi, Ran Yan, Zheng Chai, Haotian Zhang, Xiao LiuarXiv 2021 · arXiv

Abstract

The Megvii-3D team presents a robust image matching pipeline for the CVPR 2021 SimLocMatch Challenge, focusing on indoor environments. By integrating SuperPoint and DISK features with SuperGlue matchers, and employing domain-specific weights alongside a heuristic "discarding threshold," the method effectively handles repetitive textures and viewpoint changes. Experimental results demonstrate that a hybrid feature approach combined with single-scale matching yields superior performance, achieving a match success rate of 52.19%.

Introduction

Image matching—the process of finding corresponding points between two different views of a scene—is a cornerstone of computer vision tasks like Structure-from-Motion (SfM) and Visual Localization. The SimLocMatch Challenge specifically focuses on matching images in indoor environments, such as offices and restaurants. These settings are notoriously difficult due to repetitive textures, low-light conditions, and significant viewpoint changes.

A major hurdle in these challenges is not just finding matches between related images, but correctly identifying when two images do not match at all. If a system produces too many "matches" for non-matching pairs, it can lead to catastrophic failures in downstream geometric estimation. This work aims to maximize the number of correct inliers and the matching success ratio while minimizing false positives in non-matching pairs.

Methodology

The proposed approach experiments with several configurations of local features and matchers to optimize performance for indoor scenes.

Feature Extraction and Matching

The team compared using DISK (8k keypoints) against a hybrid approach combining SuperPoint (2k keypoints) and DISK (6k keypoints). Each feature type was matched using its respective SuperGlue configuration. To adapt to the indoor domain, pre-trained indoor weights for the SuperPoint-SuperGlue matcher were utilized.

Scale and Filtering Strategies

While multi-scale matching is often a standard for robustness, evaluations showed that single-scale matching yielded a higher number of inliers and a better success ratio for this specific challenge. Multi-scale matching tended to increase the number of matches for non-matching images, which is undesirable.

To handle non-matching images, a "discarding threshold" was implemented. If a pair had fewer than a certain number of matches (e.g., 8, the theoretical minimum for solving certain geometric poses), the pair was discarded. This heuristic proved effective for pruning false positive image pairs.

Outlier Rejection

DegenSAC was employed for robust geometric estimation and outlier removal without additional fine-tuning, ensuring stable pose estimation from the identified matches.

Key Results

The team submitted several variations of their pipeline, labeled sss-sd and aaa-1000k. The hybrid approach consistently outperformed single-feature baselines, suggesting that different detectors capture complementary geometric or textural information.

Table 1: Performance comparison of various configurations on the SimLocMatch dataset, highlighting inlier counts and success rates.

As shown in Table 1, the sss-sd_100k_8 configuration achieved the best results with a 52.19% match success rate and an average of 319.02 inliers. Adjusting the "discard nums" threshold allowed for significant reduction in false matches for non-matching pairs. For instance, increasing the threshold from 8 to 50 reduced non-matching "matches" from approximately 30 to 23, though this resulted in a lower overall success rate.

Conclusion

The Megvii-3D team's success in the SimLocMatch Challenge highlights the effectiveness of hybrid feature extraction and domain-specific adaptation. Leveraging pre-trained indoor weights and precise filtering thresholds are critical for navigating the complexities of indoor visual localization. Future work will focus on training DISK-SuperGlue matchers specifically for indoor scenes to further improve performance.

Summary updated

BibTeX
@misc{bi2021method,
  author = {Xiaopeng Bi and Ran Yan and Zheng Chai and Haotian Zhang and Xiao Liu},
  title = {{Method Towards CVPR 2021 SimLocMatch Challenge}},
  year = {2021},
  eprint = {2108.04466},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.04466},
}
Figure