YangParky
commented
3 months ago Hi, thanks for your interest in our work, I have reuploaded the baseg.py ~
The challenge lies in the scarcity of boundary information compared to body information. To fully utilize the available boundary data, we considering that boundary information contains long-range dependencies, thus adopt attention-based mechanism to guide context aggregation. The aggregated context is then interpolated to generate the segmentation.
Actually, after training our BASeg model, you'll see that it produces remarkably clear boundary segmentation~
Billy-ZTB
Hello! I have had the chance to read your paper and your code, your work is pretty innovative ! However, I have a question about your context aggregation module which according to your code, takes the f4 from resnet and edge from boundary stream as input. The parameter edge comming from Edge_Detection has the same height and width as f4, which has low spatial resolution. Would utilizing the low spatial resolution edge information lead to inaccuracy of the boundary of objects in images? If your input image is 256x256, the edge feature you are actually utilizing would be 33x33, it is quite blurry, I am curious about how this edge information is making contribution to the improvement of segmentaion eventually. Looking forward to your reply! 😊