Skip to main content
Back to timeline
arXivSource publication:

Two equivalent encodings of one rotated box mislead semi-supervised detectors: Oriented Semi-DETR lifts mAP at every DOTA-v2.0 label ratio

Synopsis

In aerial images a rotated box has two parameter encodings that describe the same rectangle, so a semi-supervised detector can penalize and discard a student box that is already correctly localized. Oriented Semi-DETR removes that ambiguity at the three points where a pseudo box enters training: the regression target follows whichever encoding is closer to the student prediction, the matching cost takes the minimum over both encodings, and pseudo-label denoising starts earlier and is weighted by teacher score and encoding margin. On DOTA-v2.0 at 10%, 20% and 30% labeled data, mAP exceeds the best DETR-based baseline by 1.78, 2.18 and 2.59 points, with larger gains at the strict IoU threshold.

Source-provided article image: Oriented Semi-DETR for Semi-Supervised Aerial Object Detection
arXiv

Interpretation

In semi-supervised aerial object detection, the map from rotated-box parameters to rectangles is two-to-one: one rectangle has two parameter encodings, and the regression loss, the Hungarian matching cost and the Gaussian-mixture filter all operate directly on those parameters. The earlier semi-supervised DETR methods Semi-DETR and Sparse Semi-DETR were not designed for rotated boxes and do not resolve the two equivalent encodings; at 20k iterations the swapped encoding is closer to the student prediction for 16.9% of matched pseudo-label and query pairs. On DOTA-v2.0, Oriented Semi-DETR reaches mAP of 27.61, 30.69 and 32.19 at 10%, 20% and 30% labeled data, 1.78, 2.18 and 2.59 points above the best DETR-based baseline, and 2.15, 3.39 and 4.04 points above it at the strict IoU threshold.

The method removes the ambiguity at the three places where a pseudo box enters training, adding no parameters and no extra forward pass. Swap-aligned target selection regresses toward whichever encoding is closer to the student prediction, a swap-invariant reliability cost takes the minimum over the two encodings inside the matching cost, and swap-aware pseudo-label denoising moves the denoising start from 160k to 120k with a 10k linear ramp. At 30% labeled data, applying the rule to the supervised denoising target alone raises mAP from 28.69 to 30.14, adding the ordinary regression target raises it to 31.89, and aligning all three regression targets gives 31.95.

Perspective

All experiments use DOTA-v2.0, evaluated by the official server on test-dev, with a ResNet-50 DINO detector trained for 180k iterations; the 10%, 20% and 30% splits follow Semi-DETR's partial-label protocol, and the fully labeled setting uses the entire trainval set as labeled data and the test-challenge images as unlabeled data. Every run sets the one-to-many warm-up equal to the full schedule, so the results speak to the one-to-many stage and the effect in the one-to-one stage remains untested.

The isolated effect of the reliability cost rests on an unbalanced comparison between the fixed-cost control, trained once, and the swap-invariant-cost control, trained three times, which the authors say should be rerun on equal footing. In the angle-penalty comparison the three forms take different values at the swap, so the comparison identifies a better configuration rather than isolating the shape of the function. All runs use one seed, and the reported means average out remaining nondeterminism at that seed without supporting a significance test. The diagnostic uses 128 unlabeled images, pairs from the same image are strongly correlated and no confidence intervals are given, and class composition is uneven, confounding aspect ratio with class.