🔍 TimeLens2: Precise Moment Retrieval in Video via Multimodal LLMs
A family of TimeLens2 models (4B and 8B) has been introduced, designed for precise temporal interval retrieval in videos based on text queries. The models work effectively with long videos and egocentric recordings, utilizing the GRPO method to increase accuracy.
🌍 This automates video labeling and will improve archive navigation by optimizing spatio-temporal relationships.
👤 It allows for quickly finding scenes in long videos via description, which is useful for video editing and surveillance.
Source 1: https://mcg-nju.github.io/TimeLens2/ Source 2: https://arxiv.org/abs/2607.17423