Deep Radial Embedding for Visual Sequence Learning

Connectionist Temporal Classification (CTC) is a popularobjective function in sequence recognition, which provides supervisionfor unsegmented sequence data through aligning sequence and its corresponding labeling iteratively. The blank class of CTC plays a crucialrole in the alignment process and is often considered responsible for thepeaky behavior of CTC. In this study, we propose an objective functionnamed RadialCTC that constrains sequence features on a hyperspherewhile retaining the iterative alignment mechanism of CTC. The learnedfeatures of each non-blank class are distributed on a radial arc from thecenter of the blank class, which provides a clear geometric interpretationand makes the alignment process more efficient. Besides, RadialCTC cancontrol the peaky behavior by simply modifying the logit of the blankclass. Experimental results of recognition and localization demonstratethe effectiveness of RadialCTC on two sequence recognition applications.