A Classification Refinement Strategy for Semantic Segmentation
This addresses pixel-level classification errors in semantic segmentation, though it appears incremental as it refines existing classifiers rather than introducing a fundamentally new approach.
The paper tackles semantic segmentation errors by proposing a refinement strategy that uses classifier confusion statistics and object class priors to re-estimate pixel labels. Experiments on multiple datasets show this approach can significantly improve semantic labeling performance.
Based on the observation that semantic segmentation errors are partially predictable, we propose a compact formulation using confusion statistics of the trained classifier to refine (re-estimate) the initial pixel label hypotheses. The proposed strategy is contingent upon computing the classifier confusion probabilities for a given dataset and estimating a relevant prior on the object classes present in the image to be classified. We provide a procedure to robustly estimate the confusion probabilities and explore multiple prior definitions. Experiments are shown comparing performances on multiple challenging datasets using different priors to improve a state-of-the-art semantic segmentation classifier. This study demonstrates the potential to significantly improve semantic labeling and motivates future work for reliable label prior estimation from images.