Abstract
This thesis addresses two major challenges in semantic segmentation for real-world data. First, with ever-increasing semantic labels, we need a more pragmatic approach other than existing fully-supervised methods. Second, semantic segmentation for very small or rarely-appeared objects are still very challenging for existing methods. In this thesis, we propose to (1) fully utilize the predicted label information from an existing supervised model and to (2) infer newly generated labels via label transfer from a real-world dataset. We propose a “content-adaptive” and “label-aware” MRF framework to jointly exploiting both the supervised and label-transferrable knowledge. The proposed method needs no off-line training and can easily adapt to real-world data. Experimental results on SIFT Flow and LMSun datasets demonstrate the effectiveness of the proposed method, and show promising performance over state-of-the-art methods under the real-world scenario.