Skip to main content
Back to timeline
arXivSource publication:

Testing each detection box against its own size and border stratum: PLUM flags unknown objects without training and reaches 91.5 AUROC on PASCAL-VOC

Synopsis

When a detector meets an object outside its trained classes, it still draws a box and assigns a known class with high confidence; PLUM recasts each box decision as a hypothesis test that compares the box only with known-class boxes of the same size and border contact. It reads the last backbone stage pooled inside the box and measures the distance to known-class statistics, needing no training, external model or extra forward pass. On the PASCAL-VOC and BDD100K benchmarks, PLUM has the highest AUROC in all four known/unknown pairings, at 91.5, 94.5, 95.1 and 97.5.

Source-provided article image: Box-conditional Hypothesis Testing for Out-of-Distribution Object Detection
arXiv

Interpretation

The unknown-object decision moves from a whole image or a single threshold to a per-box hypothesis test whose null hypothesis is that the box covers a known-class object. Such scores were mostly pooled over all detections of a benchmark, with one null distribution ranking every box, while detector scores change with box size and with contact with the image border. On PASCAL-VOC, changing the null distribution from one shared by all boxes to size-level strata raises AUROC from 74.2 to 91.8; adding border contact and a log-area trend line gives 93.0 with FPR95 of 36.7.

The test reads the last backbone stage pooled inside the box, normalizes that feature to unit length and measures its Mahalanobis distance to the nearest known-class mean. The feature pyramid and box head are trained only for known classes and have no reason to keep what sets an unknown object apart, so the test reads the layers before them. Across layers of Faster R-CNN, res5 has the best leveled AUROC and FPR95 in all four known/unknown pairings; past the task boundary the leveled AUROC falls clearly, for example 60.2 for fc1 and 65.0 for fc2 against 93.0 for res5 on PASCAL-VOC.

Perspective

The demonstrated setting is the standard detection benchmark with PASCAL-VOC and BDD100K as known data and MS-COCO and OpenImages as unknown data, with a ResNet-50 Faster R-CNN as the main detector and Deformable DETR results in the supplementary material. PLUM leaves detections unchanged, so known-class detection accuracy is the detector's own 48.55 and 31.24 COCO-style AP.

The false-alarm bound requires reference boxes exchangeable with test boxes, while PLUM draws references from training images and treats boxes of one image as independent. With OpenImages as unknown data and PASCAL-VOC as known data, the background-box AUROC reaches 73.5, so PLUM partly separates these images without an object.