We propose a novel form of supervision called action-agnostic point-level (AAPL) supervision
for temporal action detection.
Extensive experiments demonstrate that the proposed approach is competitive with or outperforms prior methods in terms of the trade-off between the annotation cost and detection performance.