Adaptive Template and Transition Map for Real-Time Video Object Segmentation

Semi-supervised video object segmentation (semi-VOS) is required for many visual applications. This task tracking class-agnostic objects from a given mask. Various approaches have been developed and achieved high accuracy in this field, but these previous models are hard to be utilized real-world applications due slow inference time tremendous complexity. To significantly speed up while reducing performance gaps those models, we introduce fast model based on template matching method auxiliary loss with transition map. Our consists of short-term long-term matching. The enhances target localization by focusing neighboring frames, improves fine details handles shape-changing considering long-range frames. However, since both processes generate each the previously estimated masks, incurs error propagation next mitigate problem, add newly proposed map encouraging correction power create accurate masks object. obtains $81.1\%~J\&F$ score at 78.3 FPS DAVIS16 benchmark achieves $1.4\times $ faster 11.3% higher than SiamMask, one semi-VOS models.