Anurag Das

I am a PhD student in the Computer Vision group at MPI Informatics in Saarbrücken, supervised by Prof. Bernt Schiele. I was a research intern at Samsung AI Cambridge during the summer of 2025, where I worked on multimodal large language models. Prior to that, I was part of the preparatory phase of the Saarbrücken Graduate School of Computer Science. I completed my bachelor’s degree in Information Technology at IIIT Allahabad (2013–2017). My recent interests include reasoning with multimodal LLMs and improving their efficiency.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

Although Large Vision–Language Models (LVLMs) excel on many tasks, their ability to reason over multiple images is poorly understood and insufficiently analyzed. We introduce MIMIC, a diagnostic benchmark that exposes fundamental multi-image failures in LVLMs and propose complementary data-generation and attention-masking remedies that significantly improve cross-image reasoning and achieve state-of-the-art performance on multi-image benchmarks.

We propose mask-text alignment based framework for improved semantic segmentation result. We show how mask-text alignment is better than pixel-text alignment for dense semantic segmentation task.

We propose a 3D-aware finetuning of 2D foundational models, resulting in improved performance on downstream tasks such as semantic segmentation and depth estimation.

We propose a common framework to use different weak labels, e.g., image, point and coarse labels from the target domain to reduce the performance gap between UDA and supervised learning.

We propose to utlize cheaper coarse annotations for urban scene semantic segmentation. Coarse annotation lacks fine Boundary details and are faster to annotate. Our proposed method obtains competitive performance with coarse annotation along with relatively free synthetic data compared with fine annotation at a fraction of the annotation budget.

Generalized Zero-shot Semantic Segmentation(GZSS) is a challenging problem as the prior works don't generalize well on unseen classes. we propose to leverage a class-agnostic segmentation prior provided by superpixels and introduce a superpixel pooling (SP-pooling) module that improves the performance on GZSS task.

• Reviewer: CVPR23, ECCV23, PAMI, IJCV, TMLR, IEEE-TCVST
• 2021: DAGM Best Master Thesis Award 2021 in Germany
• 2021: Saarland Stipendium funded by DAAD
• 2019: Saarbrucken Graduate School of Computer Science Scholarship
• 2017: Institute Merit Scholarship (IIIT Allahabad)
• 2011: Qualified for Indian National Mathematical Olympiad (INMO)

Stolen from Jon Barron. Big thanks!