DeepVoting: An Explainable Framework for Semantic Part Detection under Partial Occlusion

In this paper, we study the task of detecting semantic parts of an object. This is very important in computer vision, as it provides the possibility to parse an object as human do, and helps us better understand object detection algorithms. Also, detecting semantic parts is very challenging especially when the parts are partially or fully occluded. In this scenario, the popular proposal-based methods like Faster-RCNN often produce unsatisfactory results, because both the proposal extraction and classification stages may be confused by the irrelevant occluders. To this end, we propose a novel detection framework, named DeepVoting, which accumulates local visual cues, called visual concepts (VC), to locate the semantic parts. Our approach involves adding two layers after the intermediate outputs of a deep neural network. The first layer is used to extract VC responses, and the second layer performs a voting mechanism to capture the spatial relationship between VC's and semantic parts. The benefit is that each semantic part is supported by multiple VC's. Even if some of the supporting VC's are missing due to occlusion, we can still infer the presence of the target semantic part using the remaining ones. To avoid generating an exponentially large training set to cover all occlusion cases, we train our model without seeing occlusion and transfer the learned knowledge to deal with occlusions. This setting favors learning the models which are naturally robust and adaptive to occlusions instead of over-fitting the occlusion patterns in the training data. In experiments, DeepVoting shows significantly better performance on semantic part detection in occlusion scenarios, compared with Faster-RCNN, with one order of magnitude fewer parameters and 2.5x testing speed. In addition, DeepVoting is explainable as the detection result can be diagnosed via looking up the voted VC's.