Incremental and Multi-Task Learning Strategies for Coarse-To-Fine Semantic Segmentation

摘要: The semantic understanding of a scene is key problem in the computer vision field. In this work, we address multi-level segmentation task where deep neural network first trained to recognize an initial, coarse, set few classes. Then, incremental-like approach, it adapted segment and label new objects’ categories hierarchically derived from subdividing classes initial set. We propose strategies output coarse classifiers fed architectures performing finer classification. Furthermore, investigate possibility predict different levels together, which also helps achieve higher accuracy. Experimental results on New York University Depth v2 (NYUDv2) dataset show promising insights understanding.