Memory Consolidation Enables Long-Context Video Understanding

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Balažević, Ivana, Shi, Yuge, Papalampidi, Pinelopi, Chaabouni, Rahma, Koppula, Skanda, Hénaff, Olivier J.
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929367493378048
author Balažević, Ivana
Shi, Yuge
Papalampidi, Pinelopi
Chaabouni, Rahma
Koppula, Skanda
Hénaff, Olivier J.
author_facet Balažević, Ivana
Shi, Yuge
Papalampidi, Pinelopi
Chaabouni, Rahma
Koppula, Skanda
Hénaff, Olivier J.
contents Most transformer-based video encoders are limited to short temporal contexts due to their quadratic complexity. While various attempts have been made to extend this context, this has often come at the cost of both conceptual and computational complexity. We propose to instead re-purpose existing pre-trained video transformers by simply fine-tuning them to attend to memories derived non-parametrically from past activations. By leveraging redundancy reduction, our memory-consolidated vision transformer (MC-ViT) effortlessly extends its context far into the past and exhibits excellent scaling behavior when learning from longer videos. In doing so, MC-ViT sets a new state-of-the-art in long-context video understanding on EgoSchema, Perception Test, and Diving48, outperforming methods that benefit from orders of magnitude more parameters.
format Preprint
id arxiv_https___arxiv_org_abs_2402_05861
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Memory Consolidation Enables Long-Context Video Understanding
Balažević, Ivana
Shi, Yuge
Papalampidi, Pinelopi
Chaabouni, Rahma
Koppula, Skanda
Hénaff, Olivier J.
Computer Vision and Pattern Recognition
Most transformer-based video encoders are limited to short temporal contexts due to their quadratic complexity. While various attempts have been made to extend this context, this has often come at the cost of both conceptual and computational complexity. We propose to instead re-purpose existing pre-trained video transformers by simply fine-tuning them to attend to memories derived non-parametrically from past activations. By leveraging redundancy reduction, our memory-consolidated vision transformer (MC-ViT) effortlessly extends its context far into the past and exhibits excellent scaling behavior when learning from longer videos. In doing so, MC-ViT sets a new state-of-the-art in long-context video understanding on EgoSchema, Perception Test, and Diving48, outperforming methods that benefit from orders of magnitude more parameters.
title Memory Consolidation Enables Long-Context Video Understanding
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2402.05861