train_question

[hcscysh]

Your work is excellent. Figure 1 is the original RGB image, and Figures 2 and 3 are two different masks. I would like to ask, for this model, I want to train my own dataset on it, which mask would perform better?

[ZhengPeng7]

Thanks! I haven't tried BiRefNet on contour detection, but it should also be okay. Even though I still personally prefer the mask of regions if you don't need very high contours, which should be easier to train.

[hcscysh]

@ZhengPeng7 Thank you so much for your reply. Your help has generated very good results. Since we're only segmenting the wall outlines now, one image with a corresponding wall outline mask is sufficient. Now, I want to know how to create masks if I want to segment into multiple categories. I know that in YOLOv, multi-class segmentation training is based on an image's corresponding category.txt file, but in BiRefNet, how should I create my dataset? Something similar to the image below, where the color mask represents the category I want to segment.

[ZhengPeng7]

I'm not sure about the specific updates to extend BiRefNet to semantic seg. Maybe some days later, if I want to involve exps for it in the v2.