About me

I am a PhD student in the Computer Graphics and Vision group at the University of Tübingen, advised by Prof. Hendrik Lensch, and a scholar of the International Max Planck Research School for Intelligent Systems (IMPRS-IS).

My research focuses on robust video object tracking and segmentation with foundation models: keeping track of objects through occlusions and look-alike distractors, and building systems that notice when a prediction has gone wrong and correct it. I also work on self-supervised learning, histopathology image analysis and surgical video analysis.

Before Tübingen, I completed a dual degree at the Indian Institute of Technology Bombay: a B.Tech in Energy Engineering with an M.Tech in Artificial Intelligence and Data Science, and a minor in Computer Science. There I worked with Prof. Biplab Banerjee on few-shot hyperspectral image classification and domain generalization, and with Prof. Aditya Grover (UCLA) on vision-and-language navigation.

Research interests

  • Video object tracking and segmentation
  • Self-supervised learning
  • Histopathology image analysis
  • Surgical video analysis

News

  • 2026.09

    🎉 Our Dual-Agent paper is accepted at NeurIPS 2026!

  • 2026.09

    📝 Learning Where to Focus: Self-Supervised Multi-Scale ViTs for Histopathology is accepted at GCPR 2026.

  • 2026.06

    🎉 ReMeDI: Refined Memory for Disambiguation of Identities with SAM3 in Surgical Segmentation is accepted at MICCAI 2026.

  • 2025.07

    📊 Evaluating Self-Supervised Learning in Medical Imaging: A Benchmark for Robustness, Generalizability, and Multi-Domain Impact is accepted at BMVC 2025.

  • 2025.05

    🔬 Two papers accepted: HistDiST at GCPR 2025, and SurgTracker at the SynData4CV workshop, CVPR 2025.

  • 2024.04

    🧭 Scaling Vision-and-Language Navigation With Offline RL is published in TMLR.

  • 2023.11

    📄 Learning Class and Domain Augmentations for Single-Source Open-Domain Generalization is accepted at WACV 2024.

  • 2023.09

    🚀 Started my PhD at the University of Tübingen with Prof. Hendrik Lensch as an IMPRS-IS scholar.

  • 2023.07

    🎓 Graduated from IIT Bombay with a dual degree: B.Tech in Energy Engineering and M.Tech in Artificial Intelligence and Data Science.

Publications

* equal contribution

  • 2026

  • When Trackers Fail: VLM-Guided Verification and Recovery for Robust Video Object Segmentation

    NeurIPS 2026

    Valay Bundele, Susmit Agrawal, Mehran Hosseinzadeh, The Nam Nguyen, Hendrik P. A. Lensch

    A failure-aware tracking loop around SAM 3: a vision-language verification agent checks whether the propagated mask still follows the target, and a correction agent re-prompts the tracker only when it does not. Improves SAM 3 by 1.2 and 2.7 J&F on LVOS-v2 and SeCVOS.

    Project page

  • Learning Where to Focus: Self-Supervised Multi-Scale ViTs for Histopathology

    GCPR 2026

    Anabel Stammer, Valay Bundele*, Mehran Hosseinzadeh*, Hendrik P. A. Lensch

    Self-supervised multi-scale tokenization for whole-slide images: class attention decides which regions deserve fine tokens while the rest stay coarse, so capacity goes where the diagnostic signal is.

    arXiv

  • ReMeDI: Refined Memory for Disambiguation of Identities with SAM3 in Surgical Segmentation

    MICCAI 2026

    Valay Bundele, Mehran Hosseinzadeh, Hendrik P. A. Lensch

    Keeps instrument identities apart in surgical video with a refined SAM 3 memory: occlusion- and relevance-aware memory banks, disocclusion detection and feature-based re-identification, all without retraining the tracker.

    Project pagePaper

  • 2025

  • Evaluating Self-Supervised Learning in Medical Imaging: A Benchmark for Robustness, Generalizability, and Multi-Domain Impact

    BMVC 2025

    Valay Bundele*, Karahan Sarıtaş*, Bora Kargi*, Oğuz Ata Çal*, Kıvanç Tezören*, Zohreh Ghaderi, Hendrik P. A. Lensch

    A systematic study of self-supervised pretraining on medical image classification: in-domain accuracy, robustness to distribution shift and multi-domain transfer across methods, backbones and initializations.

    PaperarXiv

  • HistDiST: Histopathological Diffusion-based Stain Transfer

    GCPR 2025

    Erik Großkopf*, Valay Bundele*, Mehran Hosseinzadeh*, Hendrik P. A. Lensch

    Latent diffusion for H&E to IHC stain translation with dual conditioning on a pathology foundation model and a rescaled noise schedule, so structure is preserved while the stain changes.

    arXivCode

  • You Are Your Best Teacher: Semi-Supervised Surgical Point Tracking with Cycle-Consistent Self-Distillation

    CVPR 2025 Workshop (SynData4CV)

    Valay Bundele, Mehran Hosseinzadeh, Hendrik P. A. Lensch

    Adapts a synthetic-trained point tracker to surgical video with a frozen teacher, cycle-consistency filtering of its pseudo-labels and student fine-tuning, using only 80 unlabeled videos.

    arXiv

  • 2024

  • Scaling Vision-and-Language Navigation With Offline RL

    TMLR 2024

    Valay Bundele, Mahesh Bhupati, Biplab Banerjee, Aditya Grover

    Learning vision-and-language navigation policies from offline, suboptimal trajectory data with offline reinforcement learning.

    arXiv

  • Learning Class and Domain Augmentations for Single-Source Open-Domain Generalization

    WACV 2024

    Prathmesh Bele, Valay Bundele*, Avigyan Bhattacharya*, Ankit Jha, Gemma Roig, Biplab Banerjee

    Generalizing from a single source domain to unseen domains and unseen classes by learning class and domain augmentations jointly.

    arXivDOI

  • 2022

  • Perceptual cGAN for MRI Super-resolution

    EMBC 2022

    Sahar Almahfouz Nasser, Saqib Shamsi, Valay Bundele, Bhavesh Garg, Amit Sethi

    A conditional GAN with a perceptual loss for isotropic and anisotropic MRI super-resolution.

    arXivDOI

  • Few-Shot Open-Set Recognition of Hyperspectral Images with Outlier Calibration Network

    WACV 2022

    Debabrata Pal, Valay Bundele, Renuka Sharma, Biplab Banerjee, Yogananda Jeppu

    Recognizing new hyperspectral classes from a few examples while rejecting unknown ones, with a calibrated outlier branch.

    DOI

  • 2021

  • SPN: Stable Prototypical Network for Few-Shot Learning-Based Hyperspectral Image Classification

    IEEE GRSL 2021

    Debabrata Pal*, Valay Bundele*, Biplab Banerjee, Yogananda Jeppu

    Variance-reduced prototypes make prototypical networks stable across few-shot episodes on hyperspectral data.

    DOICode

Service and teaching

Reviewing

NeurIPS, CVPR, ECCV, WACV, BMVC, MICCAI, TMLR, Pattern Recognition

Teaching assistant

  • Practical Course: Machine Learning in Graphics, Vision and LanguageUniversity of TübingenWinter 2025/26
  • Practical Course: Vision and NLPUniversity of TübingenSummer 2024
  • Machine Learning for Remote Sensing II (head TA)IIT Bombay2022
  • Machine Learning for Remote Sensing IIIIT Bombay2021