Benchmark route
Head to head
The five images where Detecture beats every baseline by the widest margin, and the five where it loses by the widest margin, with all five partitions side by side.
Ranking is by Region Covering against the best-performing baseline on each image, not by Detecture's own score. Region Covering is used rather than mIoU because it weights every ground-truth region by area whether or not a counterpart was proposed, so a method that returns one mask cannot place well by covering half the image. Numbers under each panel are that model's Region Covering on that image; the leader is highlighted.
Region colours are per-panel indices. Red in one panel and red in the next need not be the same surface, so read the shape of the division rather than the colour pairing.
Where Detecture wins by the most
The pattern is identical across all five. LISA and Sa2VA return a single mask every time, GLaMM returns one mask or none at all, and GSVA returns five regions where the truth is two. The margin comes from returning a usable partition at all, not from a more precise boundary.
Where Detecture loses by the most
These are genuine losses, shown unedited. GSVA takes four of them and Sa2VA the fifth. Detecture returns exactly two regions on all ten images on this page, these included, so what fails here is boundary placement rather than region count: it divides the image in the wrong place. The wins split two ways. On 007625 and 004167 the baseline simply gets it right with two regions. On the other three GSVA wins with five, where scattering the image into fragments happens to cover the ground truth better than a confidently misplaced two-region split — the mirror image of the over-segmentation that costs GSVA the aggregate score.
