MotionSqueeze: Neural Motion Feature Learning for Video Understanding

TL;DR

MotionSqueeze introduces a lightweight, trainable neural module that internalizes motion feature extraction, replacing heavy optical flow computations and significantly improving video action recognition performance.

Contribution

It presents a novel neural module for internal motion feature learning, reducing computational costs while enhancing accuracy in video understanding tasks.

Findings

Outperforms state-of-the-art on Something-Something-V1&V2 datasets

Provides significant accuracy gains with minimal additional computation

Effectively replaces external optical flow methods in neural networks

Abstract

Motion plays a crucial role in understanding videos and most state-of-the-art neural models for video classification incorporate motion information typically using optical flows extracted by a separate off-the-shelf method. As the frame-by-frame optical flows require heavy computation, incorporating motion information has remained a major computational bottleneck for video understanding. In this work, we replace external and heavy computation of optical flows with internal and light-weight learning of motion features. We propose a trainable neural module, dubbed MotionSqueeze, for effective motion feature extraction. Inserted in the middle of any neural network, it learns to establish correspondences across frames and convert them into motion features, which are readily fed to the next downstream layer for better prediction. We demonstrate that the proposed method provides a significant gain on four standard benchmarks for action recognition with only a small amount of additional cost, outperforming the state of the art on Something-Something-V1&V2 datasets.