Hyderabad: The International Institute of Information Technology, Hyderabad (IIITH) has made a strong impression at the Conference on Computer Vision and Pattern Recognition (CVPR) 2025 with the acceptance of over seven research papers, including a Best Paper Award at a workshop—further consolidating its reputation in the global AI and computer vision research community.

CVPR, considered among the top conferences in computer vision, is known for its low acceptance rate of under 30% for papers and less than 5% for oral presentations. IIITH's Centre for Visual Information Technology (CVIT), which has been a regular contributor to CVPR since 2008, continued its tradition with multiple papers presented across the main conference and workshops.

One of the highlights this year was the paper VELOCITI, led by researcher Darshana Saravanan, which introduced a benchmark for testing video-language models' ability to understand compositional visual narratives. The study found that even state-of-the-art AI models like GPT-4o and Gemini-1.5-Pro performed significantly below human accuracy in reasoning about short video clips.

Saravanan also bagged the Best Paper Award at the 12th Workshop on Fine-Grained Visual Categorization for her paper on Pseudo-labelling meets Label Smoothing for Noisy Partial Label Learning, co-authored with professors Naresh Manwani and Vineet Gandhi. Their proposed algorithm, PALS, demonstrated superior performance across seven datasets, especially in fine-grained image classification tasks.

Among other notable contributions, Prof. Ravikiran Sarvadevabhatla and his team developed RoadSocial, a video question-answering dataset derived from socially-shared road event footage, aiming to improve AI's understanding of traffic scenarios. His second paper, Sketchtopia, explored AI agents learning from asynchronous multimodal interactions in the game Pictionary.

PhD scholar Aishwarya Agarwal co-authored two papers, including TIDE, which proposes training interpretable domain generalisation models capable of correcting themselves during testing, outperforming existing models by 12%.

Undergraduate researchers Vaibhav Agrawal and Haran Raajesh also had papers accepted, a rare feat at CVPR. Their work spanned controllable text-to-image generation and improved sign language translation using contextual cues.

In all, IIITH's contributions to CVPR 2025 spanned foundational models, dataset creation, and real-world AI applications, reflecting a thriving research culture that engages scholars at all levels. Faculty members credited students for pushing boundaries and creating real-world impact through their work.