Adaptive Sparse Feature Location Activation Strategy for Sparse Detectors on Drone Images

Abstract

Sparse convolution, operating convolutions on sparsely sampled areas via a learnable mask, has witnessed its powerful ability to accelerate detector inference speed on high-resolution drone images. However, the sparse mask in existing sparse convolution-based detection methods often struggles with an insufficient sampling of objects and inaccurate sampling of object locations, resulting in limited improvements in detection accuracy or even degradation. To address this problem, we introduce an adaptive Sparse Feature Location Activation (SFLA) strategy to dynamically adjust the learning objectives of sparse masks according to the current mask ratio, based on which we further design a sparse feature activation location loss which boosts the location activation accuracy of sparse feature. Extensive experimental results on the challenging UAV benchmark dataset, i.e., VisDrone, demonstrate the effectiveness of our approach.