Reproduce “Open-Vocabulary Temporal Action Localization using Multimodal Guidance” 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
Akshita Gupta, Aditya Arora, Sanath Narayan, Salman Khan, Fahad Shahbaz Khan, Graham W. Taylor
Researchers introduced Open-Vocabulary Temporal Action Localization (OVTAL) and developed OVFormer, a model that localizes and classifies any action, including those unseen during training, in untrimmed videos. OVFormer achieves 27.3% mAP on novel categories on THUMOS14 and 25.1% mAP on novel categories on ActivityNet-1.3, substantially outperforming existing baselines and setting new state-of-the-art results for zero-shot temporal action localization.
Key insights
- The explicit formulation and robust solution for Open-Vocabulary Temporal Action Localization (OVTAL) addresses a critical limitation in adaptable video understanding systems.
- Integrating LLM-generated rich semantic descriptions and progressive multimodal fusion via a Modality Mixer is crucial for effectively learning and generalizing to novel action categories.
- A two-stage training approach, involving broad pretraining and specific finetuning, enables strong performance on both base and novel categories without compromising generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.