Representation Learning via Global Temporal Alignment and Cycle-Consistency

TL;DR

This paper presents a novel weakly supervised approach for learning video representations by aligning temporal sequences using a probabilistic dynamic time warping method that enforces cycle consistency and improves various video understanding tasks.

Contribution

It introduces a probabilistic, contrastive, and differentiable dynamic time warping loss with cycle consistency for temporal sequence alignment in representation learning.

Findings

Significant improvements in action classification accuracy.

Enhanced performance in few-shot learning scenarios.

Effective video synchronization and 3D pose reconstruction.

Abstract

We introduce a weakly supervised method for representation learning based on aligning temporal sequences (e.g., videos) of the same process (e.g., human action). The main idea is to use the global temporal ordering of latent correspondences across sequence pairs as a supervisory signal. In particular, we propose a loss based on scoring the optimal sequence alignment to train an embedding network. Our loss is based on a novel probabilistic path finding view of dynamic time warping (DTW) that contains the following three key features: (i) the local path routing decisions are contrastive and differentiable, (ii) pairwise distances are cast as probabilities that are contrastive as well, and (iii) our formulation naturally admits a global cycle consistency loss that verifies correspondences. For evaluation, we consider the tasks of fine-grained action classification, few shot learning, and video synchronization. We report significant performance increases over previous methods. In addition, we report two applications of our temporal alignment framework, namely 3D pose reconstruction and fine-grained audio/visual retrieval.