Research map / Video / Video-language understanding
ForeSea: AI Forensic Search with Multi-modal Queries for Video Surveillance
Hyojin Park, Yi Li, Janghoon Cho, Sungha Choi, Jungsoo Lee, TAOTAO JING, shuai zhang, Munawar Hayat, Dashan Gao, Ning Bi, Fatih Porikli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| person-centric tracking and retrieval before the video LLM, versus feeding the full video (ablation) | VideoLLaMA3-7B · ForeSeaQA-MM | accuracy / temporal IoU: 61.6 -> 64.6 acc; 10.9 -> 12.5 IoU | Improved |
| crop frames to the tracked bounding box (ablation) | ForeSea · ForeSeaQA-MM | accuracy: 64.6 -> 61.3 | Hurt |
| add bbox coordinates as text in prompt (ablation) | ForeSea · ForeSeaQA-MM | accuracy / IoU: 64.6 -> 65.4 acc; 12.5 -> 13.8 IoU | Improved |
| add visual bbox overlay on top of coordinates (ablation) | ForeSea · ForeSeaQA-MM | accuracy / IoU: 65.4 -> 64.5 acc; 13.8 -> 12.1 IoU | No clear effect |
| increase retrieved tracks K from 3 to 5 (ablation) | ForeSea · ForeSeaQA-MM | accuracy / IoU: 65.4 -> 66.4 acc; 13.8 -> 12.1 IoU | Mixed |
| index global frames instead of person clips (ForeSea-Global) (ablation) | ForeSea · ForeSeaQA-MM | search accuracy / IoU: search 60.5 -> 57.5 acc, 17.6 -> 14.1 IoU; avg IoU 13.8 -> 16.1 | Mixed |
| combine person and global retrieval (ForeSea-Hybrid) (ablation) | ForeSea · ForeSeaQA overall | temporal IoU: 13.6 -> 18.4 | Improved |
| swap GCL encoder for SigLIP, SigLIP2 or EVA-CLIP (ablation) | ForeSea · ForeSeaQA overall | accuracy: 66.0 vs 65.5 / 65.8 / 64.8 | No clear effect |
| ForeSea framework versus VideoRAG with the same VideoLLaMA3 backbone (ablation) | VideoLLaMA3-7B · ForeSeaQA overall | temporal IoU: 9.9 -> 18.4 | Improved |
| replace video with predicted dense captions (ablation) | ForeSea · ForeSeaQA overall | accuracy: 66.0 -> 52.3 | Hurt |
| use ForeSea on open-domain VideoMME instead of VideoLLaMA3 alone | VideoLLaMA3-7B · VideoMME | accuracy: 66.2 -> 65.6 | No clear effect |
Closest papers by meaning
- ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge (NeurIPS 2026 · similarity 0.85)
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.84)
- BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation (NeurIPS 2026 · similarity 0.84)
- Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · similarity 0.84)
- MAGNET: A Multi-agent Framework for Finding Audio-Visual Needles by Reasoning over Multi-Video Haystacks (NeurIPS 2025 · similarity 0.84)
- RMIR: A Benchmark Dataset for Reasoning-Intensive Multimodal Image Retrieval (CVPR 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
