Rethinking the Misalignment Problem in Dense Object Detection

Object detection aims to localize and classify the objects in a given image, these two tasks are sensitive different object regions. Therefore, some locations predict high-quality bounding boxes but low classification scores, quite opposite. A misalignment exists between tasks, their features spatially entangled. In order solve problem, we propose plug-in Spatial-disentangled Task-aligned operator (SALT). By predicting task-aware point sets that located each task’s regions, SALT can reassign from those regions align them corresponding anchor point. for aligned disentangled. To minimize difference regression stages, Self-distillation (SDR) loss transfer knowledge refined results coarse results. On basis of SDR loss, SALT-Net, which explicitly exploits task-aligned point-set accurate Extensive experiments on MS-COCO dataset show our proposed methods consistently boost state-of-the-art dense detectors by $$\sim $$ 2 AP. Notably, SALT-Net with Res2Net-101-DCN backbone achieves 53.8 AP test-dev.