FSM: Feature Sampling Module for Object Detection

Abstract

Challenges caused by the acquisition condition of the images, the state of the objects, or the noise in the transmission of the images commonly exist in object detection. In those situations, the features of the objects extracted by CNNs contain certain uncertainty, which increases the difficulty of subsequent classification and regression. Towards enhancing the quality of the features, we propose a Feature Sampling Module (FSM), which learns multiple two-dimensional Gaussian distributions by the sampling network (SN) and applies those Gaussian masks to extract valid information of the features. With this sampling scheme, our method avoids learning the decision boundary from the low-quality features, making the overall model classification performance more robust. To ensure that the SN is capable of sampling the highest quality region, we design a novel sampling loss (SL) to measure the quality of the sampled features. Extensive experimental results validate the effectiveness of our proposed method.