In contrast to CVRL, TDL works on temporal triplets. It looks into the temporal dimension of a video and targets them as unique instances. The anchor and positive belongs to same temporal interval and has a high degree of resemblance in visual content compared to the negative.

More details can be found here.

Framework of the proposed Video-based Temporal-Discriminative Learning, which includes three steps in one iteration: Step 1) Generating temporal triplets for each video in a training batch; Step 2) Self-supervised learning with Temporal-Discriminative Loss for temporal-discriminative feature extraction with a 3D Backbone Network; Step 3) Updating Anchor Memory Bank with anchor features in each training batch. (Lines in red denote dissimilarity while lines in blue denote similarity. This figure is better viewed in color.)