AMA-Det: Enhancing Shared Head of One-Stage Object Detection With Adaptation, Merging, and Alignment

Feature adaptation(FA) and result alignment(RA) are critical issues in one-stage object detection, since FA amends feature misalignment problem by adapting sampling points to semantically significant locations, and RA corrects result misalignment problem by estimating the localization quality. For FA, the aligned and consistent prediction of sampling points is important. Previous studies directly inherit cascaded “proposal-refine” philosophy from multi-stage detectors and predict sampling points by approximating their minimal external rectangle to ground-truth bounding boxes. This manner generates poorly aligned and consistent sampling points and induces irrelevant features aggregated. Moreover, their sampling points for classification are conventionally generated through the localization branch, without feeding the corresponding features of the classification branch. For RA, previous studies have verified the superiority of utilizing the predicted edge distribution to estimate localization quality; however, their directly regressed distribution is incompatible with the cascaded regression framework. To solve these problems, we firstly propose a focused feature adaptation method by softening the supervision of the proposal points. This method can predict sampling points focused above the assigned objects with excellent alignment and consistency. Subsequently, inner-branch and cross-branch merging were investigated to promote feature sharing from the classification branch. Finally, cascaded distribution-guided result alignment is advanced and verified to predict accurate localization quality. After integrating our proposed adaptation, merging, and alignment, we created AMA-Det with an enhanced shared head, which impressively reaches 43.9 mAP with ResNet50 as the backbone. AMA-Det also achieves a 54.5 mAP by multi-scale testing on MSCOCO test-dev and outperforms all existing CNN-based one-stage counterparts.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.