Ishita Dasgupta

SVU 2025 Workshop

The First Workshop on Short-Form Video Understanding

Workshop Highlights

• • Exploring New Frontiers: This workshop brings together researchers working on understanding short-form videos across social media, marketing, business, and other domains. Short-form videos present unique challenges—rapid editing patterns, engagement-driven narratives, platform-specific formats—that differ fundamentally from traditional long-form video understanding, representing 90% of internet traffic with 2.5x higher engagement.
• • Key Topics: SV data collection and benchmarking, analysis and understanding techniques, detection of looped narratives and juxtaposed videos, and evaluation of engagement factors.
• • ICCV 2025: Recently co-organized this workshop at one of the premier computer vision conferences in October 2025, featuring invited talks from leading researchers and accepted papers on cutting-edge short-form video understanding.

Venue: ICCV 2025, Room 313B, Hawai'i Convention Center

Date: October 19, 2025, 1 PM - 5 PM HST, Honolulu, Hawai'i, USA

🎥 Short Video Understanding

My research focuses on understanding and analyzing short-form videos, which are characterized by unique editing techniques, narrative structures, and engagement patterns. Through papers like "SmartEdit" and "SKALD", I explore how models can understand editing patterns, engagement factors, and shot assembly.

This work is applied to content creation workflows, where understanding short video structures helps optimize creative decisions and improve video creation processes.

📊 Data-Driven Multimodal Understanding

I work on training models to understand multimodal data and using those insights to optimize content creation workflows. This includes developing systems that can reflect on engagement patterns and determine optimal content styles and formats based on user requests and performance data.

Through "VISIAR" and other work, I explore how vision and language can be integrated for multimodal understanding and generation, applied to content creation workflows to enable data-informed decision-making.

🤖 Dynamic Planning in Agents/MLLMs

I develop data-driven agents with advanced planning capabilities, exploring how Multimodal Large Language Models (MLLMs) can use dynamic planning to optimize workflows. This includes systems that can adaptively plan content creation strategies based on data insights and user requirements.

These agents leverage dynamic planning to create variations for different creative types, optimize content styles that resonate better with audiences, and generate content while maintaining stylistic consistency—all applied to content creation workflows.