Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Long Context Transfer from Language to Vision

About

Video sequences offer valuable temporal information, but existing large multimodal models (LMMs) fall short in understanding extremely long videos. Many works address this by reducing the number of visual tokens using visual resamplers. Alternatively, in this paper, we approach this problem from the perspective of the language model. By simply extrapolating the context length of the language backbone, we enable LMMs to comprehend orders of magnitude more visual tokens without any video training. We call this phenomenon long context transfer and carefully ablate its properties. To effectively measure LMMs' ability to generalize to long contexts in the vision modality, we develop V-NIAH (Visual Needle-In-A-Haystack), a purely synthetic long vision benchmark inspired by the language model's NIAH test. Our proposed Long Video Assistant (LongVA) can process 2000 frames or over 200K visual tokens without additional complexities. With its extended context length, LongVA achieves state-of-the-art performance on Video-MME among 7B-scale models by densely sampling more input frames. Our work is open-sourced at https://github.com/EvolvingLMMs-Lab/LongVA.

Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, Ziwei Liu• 2024

Related benchmarks

TaskDatasetResultRank
Video UnderstandingMVBench
Accuracy59.96
563
Video Question AnsweringActivityNet-QA
Accuracy50
418
Multi-discipline Multimodal UnderstandingMMMU
Accuracy42.6
363
Video UnderstandingVideoMME
Score (Overall)52.6
357
Video Question AnsweringActivityNet-QA (test)
Accuracy50
288
Long Video UnderstandingLongVideoBench
Score41.5
269
Streaming Video UnderstandingStreamingBench
Overall60
259
Visual Question AnsweringRealworldQA
Accuracy62.4
259
Spatial ReasoningVSI-Bench
Avg Score29.2
255
Video Question AnsweringVideoMME
Accuracy54.3
251
Showing 10 of 195 rows
...

Other info

Code

Follow for update