Papers › Multilevel semantic and adaptive actionness learning for weakly supervised temporal...

Multilevel semantic and adaptive actionness learning for weakly supervised temporal action localization

24 Nov 2024Neural Networks 2024 11archive 2025-07-28

Zhilin Li, Zilei Wang, Cerui Dong

Weakly supervised temporal action localization aims to identify and localize action instances in untrimmed videos with only video-level labels. Typically, most methods are based on a multiple instance learning framework that uses a top-K strategy to select salient segments to represent the entire video. Therefore fine-grained video information cannot be learned, resulting in poor action classification and localization performance. In this paper, we propose a Multilevel Semantic and Adaptive Actionness Learning Network (SAL), which is mainly composed of multilevel semantic learning (MSL) branch and adaptive actionness learning (AAL) branch. The MSL branch introduces second-order video semantics, which can capture fine-grained information in videos and improve video-level classification performance. Furthermore, we propagate second-order semantics to action segments to enhance the difference between different actions. The AAL branch uses pseudo labels to learn class-agnostic action information. It introduces a video segments mix-up strategy to enhance foreground generalization ability and adds an adaptive actionness mask to balance the quality and quantity of pseudo labels, thereby improving the stability of training. Extensive experiments show that SAL achieves state-of-the-art results on three benchmarks. Code: https://github.com/lizhilin-ustc/SAL

Code

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Results from the paper archive 2025-07-28

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections