Reproduce “HieraMamba: Video Temporal Grounding via Hierarchical Anchor-Mamba Pooling” 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
Joungbin An, Kristen Grauman
HieraMamba presents a novel hierarchical Anchor-MambaPooling architecture leveraging State-Space Models for linear-time processing of long videos, achieving precise localization of natural language queries. The model sets new state-of-the-art performance on Ego4D-NLQ, MAD, and TACoS benchmarks while demonstrating improved computational efficiency.
Key insights
- Combining linear-time State-Space Models (SSMs) with a hierarchical anchoring strategy effectively mitigates the quadratic computational cost of self-attention for long videos without losing temporal detail.
- Content-adaptive hierarchical compression through Anchor-MambaPooling, which includes interleaved global (Mamba) and local (Transformer) encoding, is more effective than naive downsampling for building multi-scale video representations.
- Complementary contrastive objectives, both self-supervised (ACC) and supervised (SPC), significantly enhance the compactness, distinctiveness, and semantic alignment of learned multi-scale video features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.