Answers are grounded in this paper and cite their sources.
Temporal action detection (TAD) localizes safety-critical procedures in untrimmed power-grid operation videos. Cross-site changes in action duration and density create level-dependent feature collapse in temporal pyramids and noisy, incomplete pseudo-labels under source-free domain adaptation. We propose Pyramid-Level Adaptive Imbalance-Aware Diversification (PLAID), which augments Mean-Teacher self-training with pyramid-level representation regularization. Duration-Aware Sampling (DAS) first constructs token matrices at duration-matched pyramid levels. Cross-Level Borrowing (CLB) then supplies temporally aligned neighboring tokens to levels with limited pseudo-segment support, stabilizing the statistics used by subsequent constraints. Channel variance and decorrelation losses preserve feature diversity, while Rank-Gap Reweighting (RGR) scales these constraints according to each level’s effective-rank decline from its running maximum. Together, these label-free components reduce reliance on imperfect pseudo-labels without source videos or target annotations. On four controlled duration and density shifts in the Power Grid Operations benchmark, PLAID improves source-only ActionFormer by 2.5–6.0 mean average precision (mAP) points and outperforms Vanilla Mean-Teacher in every setting. With TemporalMaxer, it also surpasses both baselines at every evaluated temporal intersection-over-union (tIoU) threshold in two representative shifts. The results support pyramid-aware representation recovery as a practical strategy for adapting fixed-camera power-grid detectors across sites when operational tempo and action occupancy vary and source recordings cannot be shared.
No takes yet. Share an insight, caveat, or question.
Meng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: