A Simple Recipe for Contrastively Pre-training Video-First Encoders Beyond 16 Frames

TL;DR

This paper introduces a simple, memory-efficient method for contrastively pre-training long, video-first encoders that outperform larger models on long-range video understanding benchmarks without architectural complexity.

Contribution

It demonstrates that large portions of videos can be masked during pre-training to effectively scale video encoders to longer durations, surpassing existing segment-based methods.

Findings

Masking up to 75% of videos improves scalability and performance.

Contrastive pre-training with masking outperforms larger LLM-based approaches.

The method scales to models with 1 billion parameters for long video understanding.

Abstract

Understanding long, real-world videos requires modeling of long-range visual dependencies. To this end, we explore video-first architectures, building on the common paradigm of transferring large-scale, image--text models to video via shallow temporal fusion. However, we expose two limitations to the approach: (1) decreased spatial capabilities, likely due to poor video--language alignment in standard video datasets, and (2) higher memory consumption, bottlenecking the number of frames that can be processed. To mitigate the memory bottleneck, we systematically analyze the memory/accuracy trade-off of various efficient methods: factorized attention, parameter-efficient image-to-video adaptation, input masking, and multi-resolution patchification. Surprisingly, simply masking large portions of the video (up to 75%) during contrastive pre-training proves to be one of the most robust ways to scale encoders to videos up to 4.3 minutes at 1 FPS. Our simple approach for training long video-to-text models, which scales to 1B parameters, does not add new architectural complexity and is able to outperform the popular paradigm of using much larger LLMs as an information aggregator over segment-based information on benchmarks with long-range temporal dependencies (YouCook2, EgoSchema).