Gauri Jagatap

Gauri Jagatap

Sr Multimodal AI Researcher, Dolby | PhD, New York University

Multimodal Understanding of Long-Form Video

At Dolby I work on making long-form video searchable and understandable by machines, by combining visual, audio and language signals.

Directions

  • Moment retrieval and video QA. Transformer architectures that fuse video, audio and text to find moments of interest and answer questions about long videos, and smarter frame sampling for video LLMs.
  • Event-level scene graphs. Open-world video scene graph generation that captures who did what, and when, across long content.
  • Temporal action localization. Online action localization trained from cheap point-level labels.
  • Agentic video curation. Agents that use scene graph representations to assemble personalized highlights, such as sports recaps.
  • Active speaker detection. Lightweight audio-visual models for real-time speaker detection on mobile devices.
  • Reliable vision-language models. Reducing hallucinations in large vision-language models, and predicting how audiences react to video.

Publications

VIS-W 2026

Semantic Action Graph: A Shared Representation for Agent Grounding and Human Interpretation of Sports Highlights

T. Lin, D. Chandran, G. Jagatap, C. Chen, A. Fanelli, D. Gunawan, J. Kimball

IEEE VIS Workshop on GenAI, Agents, and the Future of VIS, 2026

arXiv
ECCV 2026

OnPoint: Offline-to-Online Multi-Level Distillation for Point-Supervised Online Temporal Action Localization

S. Reza, G. Jagatap, M. Moghaddam, O. Camps, A. Fanelli

European Conference on Computer Vision, 2026

arXiv Project
ECCV 2026

Video2Reaction: Training Foundation Video Models to Predict Audience Reaction

T. Nguyen, S. Zhang, S. Shankar, G. Jagatap, D. Chandran, A. Fanelli, M. Fiterau

European Conference on Computer Vision, 2026Also at the CVPR 2026 Workshop on Emerging Directions in Data for Multimodal Foundation Models

arXiv Project
ACL 2026

Towards Mitigating Hallucinations in Large Vision-Language Models by Refining Textual Embeddings

A. Agrawal, G. KV, R. Aralikatti, G. Jagatap, J. Yuan, S. Baskar, V. Kamarshi, A. Fanelli, F. Huang

Findings of the Association for Computational Linguistics, 2026Also at the ICLR 2026 Workshop on Multimodal Intelligence

Paper arXiv
Under review

SCOUT: Segment Curation for Optimal Understanding of Long-Form Videos

A. Majee, G. Jagatap, D. Chandran, R. Iyer, A. Fanelli

CVPR-W 2025

Moment Sampling in Video LLMs for Long-Form VQA

M. Chasmai, G. Jagatap, G. KV, G. Van Horn, S. Maji, A. Fanelli

CVPR Workshop on Video Large Language Models, 2025

arXiv