Class Incremental Learning from Untrimmed Videos

Due to the diversity of actions and millions of video data uploaded daily through platforms like TikTok and Instagram, real-world deep learning methods must support untrimmed video. Therefore, we propose a more realistic and challenging scenario for class incremental learning using untrimmed videos. We encourage future work to focus on this and beat the current approaches.

ModelFrames per videoMem. Frame CapacityActivityNet-UntrimActivityNet-Trim
iCaRL41.6 × 10416.28%32.75%21.63%36.98%
iCaRL83.2 × 10416.67%31.96%21.54%33.41%
iCaRL166.4 × 10421.27%28.94%25.27%29.71%
iCaRL+TC41.6 × 10436.07%22.39%42.99%23.82%
iCaRL+TC83.2 × 10440.29%20.80%45.73%18.90%
iCaRL+TC166.4 × 10440.45%21.21%44.04%22.82%