DiagSplat · focused VLM boundary test
The VLM saw only observed images, Round-1 renders, residuals, and the 0/8/16/24 px candidates. Clean references and pseudo-GT were added only after decisions were locked. This tests the simplest possible keep-or-dilate loop.
| Frame | Selected | Hard added | R1 IoU | Selected IoU | Δ IoU |
|---|---|---|---|---|---|
| 0006 | +8 px | 1.182% | 0.762 | 0.763 | +0.000 |
| 0008 | +0 px | 0.000% | 0.835 | 0.835 | +0.000 |
| 0012 | +0 px | 0.000% | 0.757 | 0.757 | +0.000 |
| 0020 | +0 px | 0.000% | 0.555 | 0.555 | +0.000 |
| 0030 | +8 px | 1.173% | 0.662 | 0.664 | +0.003 |
VLM: dilate_once_then_pass — The Round-1 contour is slightly tight along the folding-screen outer frame and lower corners. The +8 px contour conservatively clears the full screen boundary; +16 px begins to consume visible static surroundings. Pseudo-GT IoU 0.762 → 0.763.
VLM: pass_without_dilation — The Round-1 contour already encloses the complete mirror and frame support. The +8 px candidate expands into the adjacent wall, floor, flowers, and umbrella without evidence that those regions belong to the same transient. Pseudo-GT IoU 0.835 → 0.835.
VLM: pass_without_dilation — The Round-1 soft envelope already extends beyond the visible motion-blur support. Uniform hard dilation would mostly remove surrounding static supervision rather than correct a clear semantic omission. Pseudo-GT IoU 0.757 → 0.757.
VLM: pass_without_dilation — The person, hands, mirror, hair, and lower support are already enclosed by the Round-1 contour. The +8 px candidate primarily grows into static wall and floor. Pseudo-GT IoU 0.555 → 0.555.
VLM: dilate_once_then_pass — The Round-1 contour is marginally tight around several defocused flower and stem extremities. A single +8 px step provides a small safety band; +16 px is unnecessarily broad around the background. Pseudo-GT IoU 0.662 → 0.664.
