Anurag Das

PhD Student at Max Planck Institute for Informatics

Email / CV / Google Scholar / Twitter / Github

Background

PhD StudentMax Planck Institute for Informatics

Computer Vision & Machine Learning, Saarbrücken, advised by Prof. Bernt Schiele
Research InternSamsung AI Center Cambridge

Cambridge, Fall 2025, on multimodal LLMs with Dr. Brais Martinez and Dr. Adrian Bulat
M.Sc. in Computer ScienceSaarland University

Saarbrücken Graduate School of Computer Science, Preparatory Phase
Software EngineerQualcomm

Hyderabad, 2017–19
B.Tech. in Information TechnologyIIIT Allahabad

Allahabad, 2013–17
Software Engineer InternMicrosoft

Hyderabad, Summer 2016
Student Quant ConsultantWorldQuant

Remote

Research Interests

I study how multimodal language models reason across images and views, and how to make them more reliable and efficient. My earlier work focused on semantic segmentation with limited supervision.

Publications

	OneScene-Bench: Do Vision-Language Models Integrate Evidence Across Views? Swathi Suhas, Anurag Das, Bernt Schiele, Jonas Fischer Under review, 2026 OneScene-Bench tests whether vision-language models combine evidence across views of the same scene. Across counting, consistency, and spatial reasoning tasks, twelve models struggle when cross-view correspondence is required.
	More Images, More Problems? A Controlled Analysis of VLM Failure Modes Anurag Das, Adrian Bulat, Alberto Baldrati, Ioannis Metaxas, Bernt Schiele, Georgios Tzimiropoulos, Brais Martinez ACL Findings, 2026 MIMIC exposes multi-image reasoning failures in large vision-language models. Data generation and attention masking improve cross-image reasoning and achieve state-of-the-art results on multi-image benchmarks.
	ClipTTT: CLIP-Guided Test-Time Training Helps LVLMs See Better Mriganka Nath^, Anurag Das^, Jiahao Xie, Bernt Schiele arXiv, 2026 ClipTTT uses CLIP-guided test-time adaptation on a single image to reduce hallucinations across 15 common corruptions, without changing the base model.
	Do Instance Priors Help Weakly Supervised Semantic Segmentation? Anurag Das^, Anna Kukleva, Xinting Hu, Yuki Asano, Bernt Schiele Transactions of Machine Learning Research*, 2026 SeSAM adapts SAM to weakly supervised segmentation with instance decomposition and iterative pseudo-label refinement. Paired with semi-supervised learning, it improves segmentation while reducing the need for dense annotations.
	MTA-CLIP: Language-Guided Semantic Segmentation with Mask-Text Alignment Anurag Das, Xinting Hu, Li Jiang, Bernt Schiele ECCV, 2024 MTA-CLIP aligns masks with text to improve semantic segmentation, outperforming pixel–text alignment.
	Improving 2D Feature Representations by 3D-Aware Fine-Tuning Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, Jan Eric Lenssen ECCV, 2024 3D-aware fine-tuning improves 2D model features for semantic segmentation and depth estimation.
	Weakly-Supervised Domain Adaptive Semantic Segmentation with Prototypical Contrastive Learning Anurag Das, Yongqin Xian, Dengxin Dai, Bernt Schiele CVPR, 2023 A unified framework uses image-level, point, and coarse target-domain labels to narrow the gap between unsupervised domain adaptation and supervised segmentation.
	Urban Scene Semantic Segmentation with Low-Cost Coarse Annotation Anurag Das, Yongqin Xian, Yang He, Zeynep Akata, Bernt Schiele WACV, 2023 Coarse annotations and synthetic data deliver competitive urban-scene segmentation at a fraction of the annotation budget required for fine labels.
	(SP)2Net for Generalized Zero-Label Semantic Segmentation Anurag Das, Yongqin Xian, Yang He, Zeynep Akata, Bernt Schiele GCPR, 2021 [ Best Master Thesis Award in Germany] Superpixel pooling supplies class-agnostic segmentation cues, improving generalization to unseen classes in zero-label semantic segmentation.