Silver medal Solution for Image Matching Challenge 2024

Image Matching Challenge 2024 [1] The aim of the competition is to construct 3D maps using sets of images from different scenarios, environments, and domains, such as drone shots, deep in dense forests, and nighttime shots. Compared to last year's Image Matching Challenge 2023 [2], the requirements are very high, such as model diversity, variability, and robustness. Because each photo is taken from a slightly different angle, the shadows may also vary depending on the time of day and season in which the photo was taken. One photo may have been taken from the ground, another from a step, and yet another from a drone. Matching images from different viewpoints is a fundamental computer vision problem that has not yet been fully solved. Factors such as surface texture or surroundings can cause performance degradation in an otherwise well-performing algorithm. The project developed a pipeline approach, the specific process is that the image set of each scene The image data features are first extracted using ImageNet weights from the pre-training model efficientnet-b7 [3], filtered based on the cosine distance, and the first n image pairs of the image set are sorted according to their similarity. Then the retrieved image pairs one by one use two keypoint feature detectors to extract the relevant feature point locations, use two keypoint matching algorithms to match all the matched point locations for the matching calculation, and save the matched pairs (match pairs) successfully.