Reproduce “Hierarchical Memory for Long Video QA” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yiqin Wang, Haoji Zhang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, Xiaojie Jin

The champion solution for the LOVEU Challenge @ CVPR’24, Track 1, introduces an enhanced Flash-VStream framework utilizing a hierarchical STAR Memory to efficiently compress visual tokens from long videos. This approach, augmented with automatic speech recognition, achieved first place in the challenge and substantially improved long video question answering performance on the MovieChat-1K dataset.

Key insights

  • Hierarchical memory mechanisms, such as STAR Memory, are highly effective for compressing redundant visual tokens in long videos, enabling efficient processing within memory constraints.
  • Integrating multimodal information, specifically ASR-transcribed speech, substantially improves comprehension and question-answering accuracy for narrative-rich long videos.
  • Domain-specific fine-tuning is critical for adapting pre-trained video-language models to achieve state-of-the-art performance on specific long video question-answering datasets.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.