In the baseline code, when I use the script prepare_data.py with threshold bbox_info[“visib_fract”] > 0.5 or even 0.9, I can still see some occluded objects with very small bounding boxes while visualizing. So I doubt that there is something wrong with the visibility fraction in the json files.
Could you please share some examples where you see that (e.g. scene id, image_id, object_id). Thanks!
This one for example 000000_rgb_cam1_000004.jpg (scene 0, cam1, image 000004). I use threshold of “bbox_info[“visib_fract”] > 0.8” but still see a very small bounding box on right middle edge.
Zoom-in view:
ah, got it. So the way the visibility score is computed is the number of pixels of the visible part divided the number of pixels of the whole part in that pose. So it compares the mask_visib versus mask. I.e. it only takes into account occlusions, and doesn’t take into account if the part is on the edge of the field of view and is not fully present in the frame. Hope this makes sense
1 Like