SysCV / MaskFreeVIS

Mask-Free Video Instance Segmentation [CVPR 2023]

Home Page:https://arxiv.org/abs/2303.15904

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

about the radius=5

bio-mlhui opened this issue · comments

Thanks for your time. I want to know how you set the radius=5. The get_neighbor_images_patch_color_similarity(image_t, image_t+1, kernel=3, di=3) function seems to compute patch similarity of one pixel on image_t+1 to the corresponding image_t patch with radius of 3 (9 pixels). Then, during matching, the topk_mask function computes the top 5 of these 9 pixels. But according to the Implementation details, it seems like computing top 5 of 25 pixels. (radius=5, K=5).

Did you achieve a larger radius by using dilated convolution? Or the radius in your paper only means the receptive field of one patch, and only samples 9 pixels around the region for choosing topk?

Looking forward to your reply.