Pixel-REfocused Navigated Tri-margin for Semi-supervised Action Detection

This paper identifies a novel issue, termed pixel activation-uncertainty, in semi-supervised action detection, which highlights the difficulty in distinguishing between action and background boundaries due to active motion. To address this, we propose an effective pipeline called Pixel-Refocused Navigated Tri-margin (PRENT), which adaptively leverages class-explicit knowledge. PRENT emphasizes maintaining region consistency by updating pseudo-label selection with each training epoch, ensuring continuous improvement. We introduce a class-explicit tri-margin as a soft solution to manage uncertain boundaries within latent buffer regions. This technique refines the buffer zone based on the unique characteristics of each category, thereby addressing the varying challenges in localizing actions and backgrounds. Experimental results on various benchmarks and training settings demonstrate the superiority of our method compared to state-of-the-art methods.