Scrub the footage. Check what was extracted. Compare the moving drawing.
How these views line up
The left and middle panels show the same archived 12 fps sample. The new 25 subjects use explicitly labelled filled contour proxies from legacy SAM exports; simplification and discarded components mean these are not the original binary masks. The original segmentation exports did not retain decoded source timestamps. Where a low-error, locally unique pixel match to the retained raw video passes the recorded checks, pairing uses that estimated appearance correspondence; other samples use the approximate CI time. Both clocks are shown. Appearance matching does not validate a mask. Known samples that cross into another shot are blocked from drawing pairing. Any recorded common crop is applied to source, mask and drawing alike; the source image link still opens the full sample. The cartoon seeks to its nearest recorded source time and stays paused while scrubbing. Open the comparison player for normal-speed playback. Samples outside the retained cartoon interval have no paired drawing.
The model’s IoU estimate is a model output, not a measured drawing-to-source fit. Segmentation can miss thin parts or merge several objects. These masks do not establish recognition, motion quality, or training eligibility.