Researchers from CVIT Wins the Best Paper Award at CVPR 2025

Details: By Ann Jacob; 11.Aug

Researchers from CVIT Wins the Best Paper Award at CVPR 2025

Nashville, Tennessee – June 11, 2025

A cutting-edge research paper titled "Pseudo-labelling meets Label Smoothing for Noisy Partial Label Learning" won the Best Paper Award at the 12th Workshop on Fine-Grained Visual Categorization (FGVC), a prestigious event held as part of the CVPR 2025 conference in Nashville.

Darshana Saravanan, a master's student at the Center for Visual Information Technology (CVIT), IIIT Hyderabad, whose paper, titled “Pseudo-labelling meets Label Smoothing for Noisy Partial Label Learning”, was co-authored with Dr. Naresh Manwani (ML Lab) and Dr. Vineet Gandhi (CVIT).

Ann Jacob

This recognition comes after a long and determined research journey, and we spoke to Darshana to learn more about her experiences and reflections.

The paper tackles a fundamental challenge in modern machine learning—how to effectively train models in situations where clean, perfectly labeled data is scarce, expensive to obtain, or requires domain-specific expertise.

The team’s research centers on Noisy Partial Label Learning (NPLL), a more realistic extension of Partial Label Learning (PLL). In PLL, each training instance is associated with a set of potential labels, only one of which is correct. NPLL pushes this boundary further by allowing for the possibility that even the true label might not be among the candidate labels—closer to real-world scenarios like crowd-sourced annotations where mistakes are common.

Winning the Best Paper Award at FGVC 2025 underscores both the technical novelty and the practical relevance of the work. The recognition places the research team among the leading voices in weakly supervised learning and fine-grained categorization, areas of increasing importance as the AI community seeks to build robust models with less reliance on costly annotated data.

First Time at CVPR – And a Big Win!

This year’s CVPR marked Darshana’s first experience at a top-tier international conference. “It was my first time attending a conference and it happened to be an A* conference. That made it even more special,” she shared. “I got to present the work orally in front of the FGVC community, which was an incredible experience.”

Darshana’s paper stood out for its novel approach to Noisy Partial Label Learning (NPLL) — a realistic form of weak supervision where even the true label might be missing from the provided label set. “We proposed a method that uses weighted nearest neighbor pseudo-labelling followed by training with label smoothing. It was a rigorous but rewarding journey,” she said.

Rejections and Rewrites – The Path to Success

Interestingly, the paper was not accepted the first time around. “It was rejected multiple times before finally being accepted,” Darshana admitted. “Some reviews were fair and others not so much. But I’m glad the work finally got the recognition it deserved.” One constructive piece of feedback from an earlier conference — to validate results on real-world datasets — helped them improve the quality of their submission.

From Thanjavur to IIIT-Hyderabad : Darshana’s academic journey began in Thanjavur, Tamil Nadu, where she completed her B.Tech in Information Technology from SASTRA University. She joined IIIT-Hyderabad as a master's student in August 2022, and started working with CVIT in early 2023.

“This was my first research experience. I had done coursework in machine learning, but doing original research was a completely new challenge,” she reflected. Her project was guided by Prof. Naresh Manwani, who introduced her to the problem, and Prof. Vineet Gandhi, her advisor for her master's work.

Life at CVIT – Learning Beyond the Classroom

Darshana spoke warmly about the collaborative and flexible environment at CVIT. “You can work at any time — day or night — and there’s always someone in the lab to help. The infrastructure is dependable and the learning opportunities are immense,” she shared.

She also appreciated the openness of her advisor, Prof. Vineet Gandhi. “He’s always been supportive of any new idea I wanted to try, and very open to collaborations.” During her time at CVIT, she worked on multiple projects across different domains, with three papers accepted at CVPR 2025 — two in workshops and one in the main conference, co-authored with Prof. Makarand Tapaswi and students from his lab.

Having recently completed her thesis defense, Darshana is set to begin a new chapter — as a Research Intern at Microsoft Research in Bangalore, starting this July. “It’s a continuation of my research journey,” she says, looking forward to applying her academic learning in a high-impact industry setting.

CVPR 2025 – Meeting the Global Community

Apart from her presentation, CVPR 2025 was also a chance to connect with the broader research community. Darshana joined a large group of IIIT-H alumni at a networking lunch and appreciated the chance to meet former students and faculty, including several from CVIT and related labs.

With three publications, a Best Paper award, and an exciting research internship ahead, Darshana’s journey is a testament to perseverance, collaboration, and the thriving research culture at IIIT-Hyderabad.

We congratulate her once again and wish her continued success in her research career.

If you're a current student curious about research at CVIT or looking to pursue work in machine learning and computer vision, Darshana’s journey offers a motivating blueprint to follow.