IIIT-H researchers win honours at CVPR 2026 for 3D AI advances

Hyderabad: Research on making 3D models lighter, improving how artificial intelligence understands videos, and preventing unauthorised 3D reconstruction from images featured among the papers presented by researchers from the International Institute of Information Technology, Hyderabad (IIIT-H) at the Computer Vision and Pattern Recognition (CVPR) 2026 conference and its workshops.Among the award-winning papers, Kunal Bhosikar’s “Fast and Robust Mesh Simplification for Generated and Real World 3-D Assets” was named Best Paper Runner-Up at a workshop on 3D geometry generation. The paper addresses a key challenge in 3D AI, where highly detailed models are often too heavy to store, render and process efficiently. The work proposes a mesh simplification technique that removes unnecessary triangles while preserving the model’s shape, curvature and texture, making 3D assets easier to use in areas such as medical imaging and virtual reality.

“The result is lighter 3D models that remain visually accurate but can be processed much faster,” Kunal said.Kunal also presented “PatchPoison: Poisoning Multi-View Datasets to Degrade 3D Reconstruction”, which explores a way to stop images from being turned into 3D models without consent. The paper introduces a nearly invisible digital patch that can be embedded into images and disrupt AI systems used for 3D reconstruction, producing blurred or distorted outputs.

The work points to a possible privacy tool for photographers, creators and businesses seeking to prevent unauthorised reuse of their images.Another IIIT-H paper to win honours was Darshan Singh’s “SRL-CLIP: Efficient CLIP Video Adaptation via Structured Semantic Role Labels”, which won Best Paper at a workshop on data-efficient video intelligence. The paper focuses on improving video understanding without relying on massive datasets.

Instead of training AI on simple captions, the team used structured semantic labels that capture who is doing what, and in what context.

“With this, we were able to train a strong video-understanding model using only a small amount of data, instead of millions of loosely described videos,” Darshan said.IIIT-H researchers also presented papers on occlusion-aware 3D image generation, video situation recognition, document question answering, and AI-generated descriptions of road videos tailored to different tones and contexts. The institute’s work was presented across the main conference, workshops and the newly introduced Findings track at CVPR 2026.