Prototype contrastive learning for point-supervised temporal action detection

作者：

Highlights：

• A pseudo-label generation strategy for point-supervised temporal action detection.

• Pseudo labels are efficiently updated to reduce error accumulation during training.

• Prototype contrastive constraint leads to more discriminative prototype features.

• Pseudo label generation by evaluating feature similarity in an embedding space.

• Experiments on three benchmarks demonstrate the superiority of the proposed method.

摘要

•A pseudo-label generation strategy for point-supervised temporal action detection.•Pseudo labels are efficiently updated to reduce error accumulation during training.•Prototype contrastive constraint leads to more discriminative prototype features.•Pseudo label generation by evaluating feature similarity in an embedding space.•Experiments on three benchmarks demonstrate the superiority of the proposed method.

论文关键词：Point-level supervision,Prototype learning,Contrastive learning,Pseudo-label learning,Temporal action detection

论文评审过程：Received 29 July 2022, Revised 25 September 2022, Accepted 1 October 2022, Available online 13 October 2022, Version of Record 20 October 2022.

论文官网地址：https://doi.org/10.1016/j.eswa.2022.118965