Apollo: An Exploration of Video Understanding in Large Multimodal Models

TL;DR

This paper investigates the mechanisms behind video understanding in large multimodal models, introduces key insights to improve training efficiency, and presents Apollo, a new family of models achieving state-of-the-art performance in long video perception.

Contribution

The paper uncovers transferability principles in video-LMMs, analyzes critical design choices, and introduces Apollo models that outperform existing benchmarks in video understanding tasks.

Findings

Scaling consistency allows transfer of design decisions from small to large models.

fps sampling during training significantly improves video representation.

Apollo models achieve state-of-the-art results on multiple long-video benchmarks.

Abstract

Despite the rapid integration of video perception capabilities into Large Multimodal Models (LMMs), the underlying mechanisms driving their video understanding remain poorly understood. Consequently, many design decisions in this domain are made without proper justification or analysis. The high computational cost of training and evaluating such models, coupled with limited open research, hinders the development of video-LMMs. To address this, we present a comprehensive study that helps uncover what effectively drives video understanding in LMMs. We begin by critically examining the primary contributors to the high computational requirements associated with video-LMM research and discover Scaling Consistency, wherein design and training decisions made on smaller models and datasets (up to a critical size) effectively transfer to larger models. Leveraging these insights, we explored many video-specific aspects of video-LMMs, including video sampling, architectures, data composition, training schedules, and more. For example, we demonstrated that fps sampling during training is vastly preferable to uniform frame sampling and which vision encoders are the best for video representation. Guided by these findings, we introduce Apollo, a state-of-the-art family of LMMs that achieve superior performance across different model sizes. Our models can perceive hour-long videos efficiently, with Apollo-3B outperforming most existing $7$B models with an impressive 55.1 on LongVideoBench. Apollo-7B is state-of-the-art compared to 7B LMMs with a 70.9 on MLVU, and 63.3 on Video-MME.