We are delighted to announce that 1 paper from Perceptual AI Lab (Prof. Chanho Eom) has been accepted to NeurIPS 2026.

Title: 

TF-PRVR: Training-Free Partially Relevant Video Retrieval for Real-World Generalization

Authors: 

Giyeol Kim, Chanho Eom

Abstract: 

Partially relevant video retrieval searches long, untrimmed videos for moments matching a text query. Fixed temporal partitions can blur event boundaries, while supervised retrieval systems may overfit their training domain. TF-PRVR instead uses frozen vision-language representations and frequency-based temporal analysis to build video-adaptive hierarchical segments. A multi-scale graph propagates query relevance across related segments, and moment-aware aggregation prioritizes consistent temporal evidence. The training-free approach is designed to preserve cross-domain generalization and is evaluated on datasets with varied visual and temporal characteristics.