Using segmentation to predict the absence of occluded parts

Occlusion poses a significant difficulty for detecting and localizing object keypoints and subsequent fine-grained identification. We propose a part-based face detection model that utilizes bottom-up class-specific segmentation in order to jointly detect and segment out the foreground pixels belonging to the face. The model explicitly represents occlusion of parts at the detection phase, allowing for hypothesized figure-ground segmentation to suggest coherent patterns of part occlusion. We show that this bi-directional interaction between recognition and grouping results in state-of-the-art part localization accuracy for challenging benchmarks with significant occlusion and yields substantial gains in the precision of keypoint occlusion prediction.