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.

We theoretically derive conditions under which loss functions for weak-label learning are proper and lower-bounded – two essential requirements for the class-probability estimation. This theorem leads us to a novel regularization called generalized logit squeezing, which empirically outperforms improper or unbounded losses.