A novel weakly-supervised approach for RGB-D-based nuclear waste object detection

This paper addresses the problem of RGBD-based detection and categorization of waste objects for nuclear decommissioning. To enable autonomous robotic manipulation for nuclear decommissioning, nuclear waste objects must be detected and categorized. However, as a novel industrial application, large amounts of annotated waste object data are currently unavailable. To overcome this problem, we propose a weakly supervised learning approach which is able to learn a deep convolutional neural network from unlabeled RGBD videos while requiring very few annotations. The proposed method also has the potential to be applied to other household or industrial applications. We evaluate our approach on the Washington RGB-D object recognition benchmark, achieving the state-of-the-art performance among semi-supervised methods. More importantly, we introduce a novel dataset, i.e., Birmingham nuclear waste simulants dataset, and evaluate our proposed approach on this novel industrial object recognition challenge. We further propose a complete real-time pipeline for RGBD-based detection and categorization of nuclear waste simulants. Our weakly supervised approach has demonstrated to be highly effective in solving a novel RGB-D object detection and recognition application with limited human annotations.