segment-anything
segment-anything copied to clipboard
facebookresearch
bad results with multiple point prompts
The segmentation looks perfect for the first point. But the result looks not reasonable if added more points as prompts. But the demo website seems work better. Is there any "tricks" with the prompts? Or mask_input of SamPredictor?
The first piont always gives perfect result, because this point is set under preview mode, which avoids confusing position. If the online demo support previewing for every single point, I think we can get perfect cuts for all objects.
I am experiencing the same issue. Using a single foreground point gives good results, however, adding background points messes everything up to the point where results are unusable. In the demo gui, negative clicks won't have the same effect.
One foreground click (quite accurate but doesn't handle overlaps well in my data with some objects):
One foreground + multiple background clicks (distributed on remaining objects, some of which are quite small):
Point labels are passed correctly as 1 for foreground and 0 for background clicks (int). I pass all clicks simultaneously with multimask_output=False. Doing a single positive click first and then additionally passing the mask within a second iteration changes results but they're still very much unusable.
Same, I have tried to reduce the number of background points, and to some extent, it helps, but I still got unsatisfied results.
Same issue.
I tried randomly sampling points from the foreground as point prompt outputs.
The results of 200 sampled points look worse than those of 10 sampled points.
