Image Matching Challenge#
Image Matching Challenge 2024 - Hexathlon
This competition aims to utilize computer vision technology to match the same scene from different perspectives, thereby recovering the camera's intrinsic and extrinsic parameter matrices at the time each image was captured. This year's competition primarily focuses on outdoor scenes and transparent low-texture objects, with a particular emphasis on matching strategies for outdoor scenes to summarize the approaches of the top contenders.
First Place: High Image Resolution ALIKED/LightGlue + Transparent Trick#
In general, the approach involves initially performing a sparse matching to align some rotated images, followed by matching the images using ALIKED and LightGlue, and employing DBSCAN to obtain dense matching regions. A second round of dense matching is then conducted, with the matching points input into COLMAP for pose estimation. The first-place team completely avoided using dense-based methods, believing that the matching points obtained through such methods typically exist only between two images, lacking transitivity, which can lead to errors in model computation.
Highlights:#
- Determining matching pairs based on sparse image matching
- Fine-tuned ALIKED and LightGlue
- Various optimizations for operational efficiency, such as caching and distributed training
- Feature point matching at multiple scales, matching once at 1280 resolution, then using DBSCAN clustering to obtain dense matching regions, cropping the images, and performing another match at 2048 resolution for finer matching
Second Place: MST-Aided SfM & Transparent Scene Solution#
- Conducted rotation detection and transparency detection on the images to separately process transparent and non-transparent images (most methods handle these two types of images separately);
- Performed global feature extraction on all images to generate more reliable image matching pairs;
- Used three methods (Dedodev2+Dual SoftMax), (DISK+LightGlue), (SIFT+NN) to obtain matching points based on the image matching pairs;
- Employed a minimum spanning tree-based strategy to obtain globally optimal adjacent graphs for the first rough SfM;
- Simultaneously used the globally optimal adjacent graph and fully connected graph for SfM, obtaining two SfM results, using the rough SfM result to filter out errors in the fine SfM result, and then utilized the relocation modules in PixSFM and HLoc to handle missing cameras.
- No further details on the processing of transparent images.
Highlights:#
- Primarily focused on using MST to obtain globally optimal graphs and using rough SfM results to filter out errors in fine SfM results