FlowNetU: Accurate Uncertainty Estimation of Optical Flow for Video Object Detection

Abstract

Video object detection (VOD) is a challenging task to resolve ambiguities owing to various issues such as motion blur and occlusion. Although various types of ambiguities will take place per pixels in an image, flow fields make equal contributions for VOD across the image. This may increase false positive (FP) results. In this paper, we propose a method that utilizes motion uncertainty for VOD. The trained optical flow estimation model helps detector to suppress unreliable flow fields in order to avoid misaggregation which causes mislocalization. Our proposed method improves mean average precision by 1.27% and decreases the FP rate by 10.59%. This verifies that utilizing motion uncertainty for video recognition tasks is very effective.