alcinos

Note that I tested it on the latest version available on git, with python2 and python3, with in both cases the same incorrect output

Apparently, the issue comes from an incorrect interpretation of the position relative/absolute in td cells. Here is a smaller example: https://gist.github.com/alcinos/140dadd8179a0811f4fe Firefox renders the previous file as follows:  Weasy...

As a side note, the rendering is correct when the table is removed and the td replaced by a regular div

It's quite hard to tell what's happening without looking at the image. One thing you can play with is the detection threshold. It's 0.5 by default, you can try lowering...

Could you provide the clean image, the point you selected, and the code you used?

What you are writing is partially true, but incomplete: it's fairly easy to write some heuristics to generate "truly semantic segmentation masks" from SAM3's outputs (and that's what we did...

Could you let us know what questions you still have after reading https://github.com/facebookresearch/sam3/blob/main/README_TRAIN.md?

Roboflow100 and odinW are not part of the training mix of the released checkpoints. We report in the paper the few-shot results on those datasets (in that case, there is...

Some of the analysis is not quite correct, for example you cite nms as a blocker, while it has an explicit cpu fallback: https://github.com/facebookresearch/sam3/blob/84cc43bca4347b772f17d1078a1ddb4c054655c2/sam3/perflib/nms.py#L63-L71 On a more general note, I...

@quangkmhd @csqqlee The error "ValueError: matrix contains invalid numeric entries" means your training is diverging. You can try the usual mitigations: bigger batch size, lower learning rate. @MinGiSa It looks...