Anurag Das (PhD Student)

MSc Anurag Das

Address: Max-Planck-Institut für Informatik
Saarland Informatics Campus
Campus E1 4
66123 Saarbrücken
Location: E1 4 - 620
Phone: +49 681 9325 2134
Fax: +49 681 9325 2099
E-mail: Get email via email

Personal Information

Personal website : https://anurag-198.github.io/

My research interests in general lies in learning with limited labels for vision tasks. I am looking for potential hiwi/master thesis students to work on related areas such as domain adaptation, few-shot, open world setting etc for semantic segmentation task. Please reach out to me if you are interested in a collaboration.

Publications

2026

Conference paper

A. Das, A. Bulat, A. Baldrati, I. M. Metaxas, B. Schiele, G. Tzimiropoulos, and B. Martinez

“More Images, More Problems? A Controlled Analysis of VLM Failure Modes,” in Findings of the Association for Computational Linguistics (ACL 2026), San Diego, CA, USA, 2026.

@inproceedings{Das_ACL26,
TITLE = {More Images, More Problems? {A} Controlled Analysis of {VLM} Failure Modes},
AUTHOR = {Das, Anurag and Bulat, Adrian and Baldrati, Alberto and Metaxas, Ioannis Maniadis and Schiele, Bernt and Tzimiropoulos, Georgios and Martinez, Brais},
LANGUAGE = {eng},
ISBN = {979-8-89176-395-1},
DOI = {10.18653/v1/2026.findings-acl.366},
PUBLISHER = {ACL},
YEAR = {2026},
DATE = {2026},
BOOKTITLE = {Findings of the Association for Computational Linguistics (ACL 2026)},
EDITOR = {Liakata, Maria and Moreira, Viviane P. and Zhang, Jiajun and Jurgens, David},
PAGES = {7423--7442},
ADDRESS = {San Diego, CA, USA},
}

Endnote

%0 Conference Proceedings
%A Das, Anurag
%A Bulat, Adrian
%A Baldrati, Alberto
%A Metaxas, Ioannis Maniadis
%A Schiele, Bernt
%A Tzimiropoulos, Georgios
%A Martinez, Brais
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
External Organizations
External Organizations
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
External Organizations
%T More Images, More Problems? A Controlled Analysis of VLM Failure Modes : 
%G eng
%U http://hdl.handle.net/21.11116/0000-0012-6EEA-A
%R 10.18653/v1/2026.findings-acl.366
%D 2026
%B 64th Annual Meeting of the Association for Computational Linguistics
%Z date of event: 2026-07-02 - 2026-07-07
%C San Diego, CA, USA
%B Findings of the Association for Computational Linguistics
%E Liakata, Maria; Moreira, Viviane P.; Zhang, Jiajun; Jurgens, David
%P 7423 - 7442
%I ACL
%@ 979-8-89176-395-1

Paper

A. Das, A. Kukleva, X. Hu, Y. M. Asano, and B. Schiele

“Do Instance Priors Help Weakly Supervised Semantic Segmentation?,” 2026. [Online]. Available: https://arxiv.org/abs/2604.11170.

Abstract

Semantic segmentation requires dense pixel-level annotations, which are costly and time-consuming to acquire. To address this, we present SeSAM, a framework that uses a foundational segmentation model, i.e. Segment Anything Model (SAM), with weak labels, including coarse masks, scribbles, and points. SAM, originally designed for instance-based segmentation, cannot be directly used for semantic segmentation tasks. In this work, we identify specific challenges faced by SAM and determine appropriate components to adapt it for class-based segmentation using weak labels. Specifically, SeSAM decomposes class masks into connected components, samples point prompts along object skeletons, selects SAM masks using weak-label coverage, and iteratively refines labels using pseudo-labels, enabling SAM-generated masks to be effectively used for semantic segmentation. Integrated with a semi-supervised learning framework, SeSAM balances ground-truth labels, SAM-based pseudo-labels, and high-confidence pseudo-labels, significantly improving segmentation quality. Extensive experiments across multiple benchmarks and weak annotation types show that SeSAM consistently outperforms weakly supervised baselines while substantially reducing annotation cost relative to fine supervision.

BibTeX

@online{Das_2604.11170,
TITLE = {Do Instance Priors Help Weakly Supervised Semantic Segmentation?},
AUTHOR = {Das, Anurag and Kukleva, Anna and Hu, Xinting and Asano, Yuki M. and Schiele, Bernt},
LANGUAGE = {eng},
URL = {https://arxiv.org/abs/2604.11170},
EPRINT = {2604.11170},
EPRINTTYPE = {arXiv},
YEAR = {2026},
ABSTRACT = {Semantic segmentation requires dense pixel-level annotations, which are costly and time-consuming to acquire. To address this, we present SeSAM, a framework that uses a foundational segmentation model, i.e. Segment Anything Model (SAM), with weak labels, including coarse masks, scribbles, and points. SAM, originally designed for instance-based segmentation, cannot be directly used for semantic segmentation tasks. In this work, we identify specific challenges faced by SAM and determine appropriate components to adapt it for class-based segmentation using weak labels. Specifically, SeSAM decomposes class masks into connected components, samples point prompts along object skeletons, selects SAM masks using weak-label coverage, and iteratively refines labels using pseudo-labels, enabling SAM-generated masks to be effectively used for semantic segmentation. Integrated with a semi-supervised learning framework, SeSAM balances ground-truth labels, SAM-based pseudo-labels, and high-confidence pseudo-labels, significantly improving segmentation quality. Extensive experiments across multiple benchmarks and weak annotation types show that SeSAM consistently outperforms weakly supervised baselines while substantially reducing annotation cost relative to fine supervision.},
}

Endnote

%0 Report
%A Das, Anurag
%A Kukleva, Anna
%A Hu, Xinting
%A Asano, Yuki M.
%A Schiele, Bernt
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
%T Do Instance Priors Help Weakly Supervised Semantic Segmentation? :
%G eng
%U http://hdl.handle.net/21.11116/0000-0012-F7F8-E
%U https://arxiv.org/abs/2604.11170
%D 2026
%X Semantic segmentation requires dense pixel-level annotations, which are costly and time-consuming to acquire. To address this, we present SeSAM, a framework that uses a foundational segmentation model, i.e. Segment Anything Model (SAM), with weak labels, including coarse masks, scribbles, and points. SAM, originally designed for instance-based segmentation, cannot be directly used for semantic segmentation tasks. In this work, we identify specific challenges faced by SAM and determine appropriate components to adapt it for class-based segmentation using weak labels. Specifically, SeSAM decomposes class masks into connected components, samples point prompts along object skeletons, selects SAM masks using weak-label coverage, and iteratively refines labels using pseudo-labels, enabling SAM-generated masks to be effectively used for semantic segmentation. Integrated with a semi-supervised learning framework, SeSAM balances ground-truth labels, SAM-based pseudo-labels, and high-confidence pseudo-labels, significantly improving segmentation quality. Extensive experiments across multiple benchmarks and weak annotation types show that SeSAM consistently outperforms weakly supervised baselines while substantially reducing annotation cost relative to fine supervision.
%K Computer Science, Computer Vision and Pattern Recognition, cs.CV

Paper

M. Nath, A. Das, J. Xie, and B. Schiele

“ClipTTT: CLIP-Guided Test-Time Training Helps LVLMs See Better,” 2026. [Online]. Available: https://arxiv.org/abs/2603.26486.

Abstract

Large vision-language models (LVLMs) tend to hallucinate, especially when visual inputs are corrupted at test time. We show that such corruptions act as additional distribution shifts, significantly amplifying hallucination rates in real-world applications. To address this, we propose CLIP-guided Test-Time Training (ClipTTT), a method to adapt LVLMs under degraded conditions on the fly with a single test sample. Specifically, we leverage the image-text alignment strength of a pre-trained CLIP model as a stable guidance signal to identify reliable self-supervision targets, enabling rapid adaptation without altering the base LVLMs. Extensive experiments on standard hallucination benchmarks, with 15 common corruptions, demonstrate that ClipTTT effectively mitigates hallucinations and improves descriptive faithfulness under visual corruptions.

BibTeX

@online{Nath_2603.26486,
TITLE = {{ClipTTT}: {CLIP}-Guided Test-Time Training Helps {LVLMs} See Better},
AUTHOR = {Nath, Mriganka and Das, Anurag and Xie, Jiahao and Schiele, Bernt},
LANGUAGE = {eng},
URL = {https://arxiv.org/abs/2603.26486},
EPRINT = {2603.26486},
EPRINTTYPE = {arXiv},
YEAR = {2026},
ABSTRACT = {Large vision-language models (LVLMs) tend to hallucinate, especially when visual inputs are corrupted at test time. We show that such corruptions act as additional distribution shifts, significantly amplifying hallucination rates in real-world applications. To address this, we propose CLIP-guided Test-Time Training (ClipTTT), a method to adapt LVLMs under degraded conditions on the fly with a single test sample. Specifically, we leverage the image-text alignment strength of a pre-trained CLIP model as a stable guidance signal to identify reliable self-supervision targets, enabling rapid adaptation without altering the base LVLMs. Extensive experiments on standard hallucination benchmarks, with 15 common corruptions, demonstrate that ClipTTT effectively mitigates hallucinations and improves descriptive faithfulness under visual corruptions.},
}

Endnote

%0 Report
%A Nath, Mriganka
%A Das, Anurag
%A Xie, Jiahao
%A Schiele, Bernt
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
%T ClipTTT: CLIP-Guided Test-Time Training Helps LVLMs See Better : 
%G eng
%U http://hdl.handle.net/21.11116/0000-0012-F836-8
%U https://arxiv.org/abs/2603.26486
%D 2026
%X Large vision-language models (LVLMs) tend to hallucinate, especially when visual inputs are corrupted at test time. We show that such corruptions act as additional distribution shifts, significantly amplifying hallucination rates in real-world applications. To address this, we propose CLIP-guided Test-Time Training (ClipTTT), a method to adapt LVLMs under degraded conditions on the fly with a single test sample. Specifically, we leverage the image-text alignment strength of a pre-trained CLIP model as a stable guidance signal to identify reliable self-supervision targets, enabling rapid adaptation without altering the base LVLMs. Extensive experiments on standard hallucination benchmarks, with 15 common corruptions, demonstrate that ClipTTT effectively mitigates hallucinations and improves descriptive faithfulness under visual corruptions.
%K Computer Science, Computer Vision and Pattern Recognition, cs.CV

2024

Conference paper

A. Das, X. Hu, L. Jiang, and B. Schiele

“MTA-CLIP: Language-Guided Semantic Segmentation with Mask-Text Alignment,” in Computer Vision -- ECCV 2024, Milano, Italy, 2024.

@inproceedings{DasECCV24,
TITLE = {{MTA-CLIP}: {L}anguage-Guided Semantic Segmentation with Mask-Text Alignment},
AUTHOR = {Das, Anurag and Hu, Xinting and Jiang, Li and Schiele, Bernt},
LANGUAGE = {eng},
ISBN = {978-3-031-72948-5},
DOI = {10.1007/978-3-031-72949-2_3},
PUBLISHER = {Springer},
YEAR = {2024},
MARGINALMARK = {$\bullet$},
DATE = {2024},
BOOKTITLE = {Computer Vision -- ECCV 2024},
EDITOR = {Leonardis, Ale{\v s} and Ricci, Elisa and Roth, Stefan and Russakovsky, Olga and Sattler, Torsten and Varol, G{\"u}l},
PAGES = {39--56},
SERIES = {Lecture Notes in Computer Science},
VOLUME = {15112},
ADDRESS = {Milano, Italy},
}

Endnote

%0 Conference Proceedings
%A Das, Anurag
%A Hu, Xinting
%A Jiang, Li
%A Schiele, Bernt
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
%T MTA-CLIP: Language-Guided Semantic Segmentation with Mask-Text Alignment : 
%G eng
%U http://hdl.handle.net/21.11116/0000-000F-D869-8
%R 10.1007/978-3-031-72949-2_3
%D 2024
%B 18th European Conference on Computer Vision 
%Z date of event: 2024-09-29 - 2024-10-04
%C Milano, Italy
%B Computer Vision -- ECCV 2024
%E Leonardis, Ale&#353;; Ricci, Elisa; Roth, Stefan; Russakovsky, Olga; Sattler, Torsten; Varol, G&#252;l
%P 39 - 56
%I Springer
%@ 978-3-031-72948-5
%B Lecture Notes in Computer Science
%N 15112

Conference paper

Y. Yue, A. Das, F. Engelmann, S. Tang, and J. E. Lenssen

“Improving 2D Feature Representations by 3D-Aware Fine-Tuning,” in Computer Vision -- ECCV 2024, Milano, Italy, 2024.

@inproceedings{YueECCV24,
TITLE = {Improving {2D} Feature Representations by {3D}-Aware Fine-Tuning},
AUTHOR = {Yue, Yuanwen and Das, Anurag and Engelmann, Francis and Tang, Siyu and Lenssen, Jan Eric},
LANGUAGE = {eng},
ISBN = {978-3-031-72627-9},
DOI = {10.1007/978-3-031-72627-9_4},
PUBLISHER = {Springer},
YEAR = {2024},
MARGINALMARK = {$\bullet$},
DATE = {2024},
BOOKTITLE = {Computer Vision -- ECCV 2024},
EDITOR = {Leonardis, Ale{\v s} and Ricci, Elisa and Roth, Stefan and Russakovsky, Olga and Sattler, Torsten and Varol, G{\"u}l},
PAGES = {57--74},
SERIES = {Lecture Notes in Computer Science},
VOLUME = {15060},
ADDRESS = {Milano, Italy},
}

Endnote

%0 Conference Proceedings
%A Yue, Yuanwen
%A Das, Anurag
%A Engelmann, Francis
%A Tang, Siyu
%A Lenssen, Jan Eric
%+ External Organizations
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
External Organizations
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
%T Improving 2D Feature Representations by 3D-Aware Fine-Tuning : 
%G eng
%U http://hdl.handle.net/21.11116/0000-000F-D86C-5
%R 10.1007/978-3-031-72627-9_4
%D 2024
%B 18th European Conference on Computer Vision 
%Z date of event: 2024-09-29 - 2024-10-04
%C Milano, Italy
%B Computer Vision -- ECCV 2024
%E Leonardis, Ale&#353;; Ricci, Elisa; Roth, Stefan; Russakovsky, Olga; Sattler, Torsten; Varol, G&#252;l
%P 57 - 74
%I Springer
%@ 978-3-031-72627-9
%B Lecture Notes in Computer Science
%N 15060

2023

Conference paper

A. Das, Y. Xian, D. Dai, and B. Schiele

“Weakly-Supervised Domain Adaptive Semantic Segmentation With Prototypical Contrastive Learning,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2023), Vancouver, Canada, 2023.

@inproceedings{Das_CVPR23,
TITLE = {Weakly-Supervised Domain Adaptive Semantic Segmentation With Prototypical Contrastive Learning},
AUTHOR = {Das, Anurag and Xian, Yongqin and Dai, Dengxin and Schiele, Bernt},
LANGUAGE = {eng},
ISBN = {979-8-3503-0129-8},
DOI = {10.1109/CVPR52729.2023.01481},
PUBLISHER = {IEEE},
YEAR = {2023},
MARGINALMARK = {$\bullet$},
DATE = {2023},
BOOKTITLE = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2023)},
PAGES = {15434--15443},
ADDRESS = {Vancouver, Canada},
}

Endnote

%0 Conference Proceedings
%A Das, Anurag
%A Xian, Yongqin
%A Dai, Dengxin
%A Schiele, Bernt
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
%T Weakly-Supervised Domain Adaptive Semantic Segmentation With Prototypical Contrastive Learning : 
%G eng
%U http://hdl.handle.net/21.11116/0000-000D-1F74-F
%R 10.1109/CVPR52729.2023.01481
%D 2023
%B 36th IEEE/CVF Conference on Computer Vision and Pattern Recognition
%Z date of event: 2023-06-18 - 2023-06-23
%C Vancouver, Canada
%B IEEE/CVF Conference on Computer Vision and Pattern Recognition
%P 15434 - 15443
%I IEEE
%@ 979-8-3503-0129-8

Conference paper

A. Das, Y. Xian, Y. He, Z. Akata, and B. Schiele

“Urban Scene Semantic Segmentation With Low-Cost Coarse Annotation,” in 2023 IEEE Winter Conference on Applications of Computer Vision (WACV 2023), Waikoloa, HI, USA, 2023.

@inproceedings{Das_WACV23,
TITLE = {Urban Scene Semantic Segmentation With Low-Cost Coarse Annotation},
AUTHOR = {Das, Anurag and Xian, Yongqin and He, Yang and Akata, Zeynep and Schiele, Bernt},
LANGUAGE = {eng},
ISBN = {978-1-6654-9346-8},
DOI = {10.1109/WACV56688.2023.00592},
PUBLISHER = {IEEE},
YEAR = {2023},
MARGINALMARK = {$\bullet$},
BOOKTITLE = {2023 IEEE Winter Conference on Applications of Computer Vision (WACV 2023)},
PAGES = {5967--5976},
ADDRESS = {Waikoloa, HI, USA},
}

Endnote

%0 Conference Proceedings
%A Das, Anurag
%A Xian, Yongqin
%A He, Yang
%A Akata, Zeynep
%A Schiele, Bernt
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
External Organizations
External Organizations
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
%T Urban Scene Semantic Segmentation With Low-Cost Coarse Annotation  : 
%G eng
%U http://hdl.handle.net/21.11116/0000-000C-6D3C-8
%R 10.1109/WACV56688.2023.00592
%D 2023
%B IEEE Winter Conference on Applications of Computer Vision
%Z date of event: 2023-01-03 - 2023-01-07
%C Waikoloa, HI, USA
%B 2023 IEEE Winter Conference on Applications of Computer Vision
%P 5967 - 5976
%I IEEE
%@ 978-1-6654-9346-8

2021

Conference paper

A. Das, Y. Xian, Y. He, B. Schiele, and Z. Akata

“(SP)2Net for Generalized Zero-Label Semantic Segmentation,” in Pattern Recognition (GCPR 2021), Bonn, Germany, 2022.

@inproceedings{Das_GCPR2021,
TITLE = {{SP$^2$Net} for Generalized Zero-Label Semantic Segmentation},
AUTHOR = {Das, Anurag and Xian, Yongqin and He, Yang and Schiele, Bernt and Akata, Zeynep},
LANGUAGE = {eng},
ISBN = {978-3-030-92659-5; 978-3-030-92658-8},
DOI = {10.1007/978-3-030-92659-5_15},
PUBLISHER = {Springer},
YEAR = {2021},
BOOKTITLE = {Pattern Recognition (GCPR 2021)},
EDITOR = {Bauckhage, Christian and Gall, J{\"u}rgen and Schwing, Alexander},
PAGES = {235--249},
SERIES = {Lecture Notes in Computer Science},
VOLUME = {13024},
ADDRESS = {Bonn, Germany},
}

Endnote

%0 Conference Proceedings
%A Das, Anurag
%A Xian, Yongqin
%A He, Yang
%A Schiele, Bernt
%A Akata, Zeynep
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
External Organizations
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
External Organizations
%T (SP)2Net for Generalized Zero-Label Semantic Segmentation : 
%G eng
%U http://hdl.handle.net/21.11116/0000-000A-6FA0-5
%R 10.1007/978-3-030-92659-5_15
%D 2022
%B 43rd German Conference on Pattern Recognition
%Z date of event: 2021-09-28 - 2021-10-01
%C Bonn, Germany
%B Pattern Recognition
%E Bauckhage, Christian; Gall, J&#252;rgen; Schwing, Alexander
%P 235 - 249
%I Springer
%@ 978-3-030-92659-5 978-3-030-92658-8
%B Lecture Notes in Computer Science
%N 13024
%U https://rdcu.be/cNoit

Thesis

A. Das

“(SP)2Net for Generalized Zero-Label Semantic Segmentation,” Universität des Saarlandes, Saarbrücken, 2021.

@mastersthesis{Das_Master21,
TITLE = {{SP$^2$Net} for Generalized Zero-Label Semantic Segmentation},
AUTHOR = {Das, Anurag},
LANGUAGE = {eng},
SCHOOL = {Universit{\"a}t des Saarlandes},
ADDRESS = {Saarbr{\"u}cken},
YEAR = {2021},
DATE = {2021},
}

Endnote

%0 Thesis
%A Das, Anurag
%Y Schiele, Bernt
%Y Xian, Yongqin
%Y He, Yang
%+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society
%T (SP)2Net for Generalized Zero-Label Semantic Segmentation : 
%G eng
%U http://hdl.handle.net/21.11116/0000-000A-A76F-E
%I Universit&#228;t des Saarlandes
%C Saarbr&#252;cken
%D 2021
%P 63 p.
%V master
%9 master