Revisiting Feature Alignment for One-stage Object Detection

TL;DR

This paper identifies a misalignment issue in one-stage object detectors and proposes a new operator and architecture, AlignDet, that improves detection accuracy by aligning features and anchors more effectively.

Contribution

The paper introduces the RoIConv operator and the AlignDet architecture, providing a principled solution to feature-anchor misalignment in one-stage detectors.

Findings

Achieves 44.1 mAP on COCO test-dev with ResNeXt-101.

Reveals the connection between im2col and RoIAlign operators.

Demonstrates improved detection performance over existing methods.

Abstract

Recently, one-stage object detectors gain much attention due to their simplicity in practice. Its fully convolutional nature greatly reduces the difficulty of training and deployment compared with two-stage detectors which require NMS and sorting for the proposal stage. However, a fundamental issue lies in all one-stage detectors is the misalignment between anchor boxes and convolutional features, which significantly hinders the performance of one-stage detectors. In this work, we first reveal the deep connection between the widely used im2col operator and the RoIAlign operator. Guided by this illuminating observation, we propose a RoIConv operator which aligns the features and its corresponding anchors in one-stage detection in a principled way. We then design a fully convolutional AlignDet architecture which combines the flexibility of learned anchors and the preciseness of aligned features. Specifically, our AlignDet achieves a state-of-the-art mAP of 44.1 on the COCO test-dev with ResNeXt-101 backbone.