Open lsylsy0516 opened 1 month ago
Hi, @lsylsy0516 Thanks for your question and thanks for trying our model on your datasets! Actually, we don't specifically try our method on small boxes with large images. Most of the images contained in our benchmark have regular boxes. Thus, it is possible that our model drops upon this situation. However, from our experimental practices, I think there are still several things you can try: 1) rerun the prototype feature extraction and make sure the prototype features are well extracted; 2) try multiple inferences with images or patches of different scales as input. The challenges of cross-domain and small box are somewhat different, but still, I hope you can get good results. There are also some papers about small box object detection, maybe could also combine their practice and see if they could combined with ours.
Hello, thanks for the great work.
I am experiencing an issue with my test dataset where the performance is excellent when the images are smaller and the bounding boxes are larger. However, when I use 1920x1080 images and the annotated bbox dimensions are within 40x30 size , the performance drops significantly.
What parameters should I adjust to address this issue?