Ayush Lodh (আযুষ)

I am a Master’s student in the Department of Computer Science & Engineering at National Institute of Technology Delhi, working under the supervision of Dr. Nisha Singh Chauhan and co-supervised by Dr. Sanket Biswas (Computer Vision Center, UAB, Barcelona).

My research spans computer vision, deep learning, and pattern recognition, with a particular focus on handwriting recognition and document understanding. My recent work includes:

From Chunks to Graphs — a training-free multimodal retrieval framework that replaces chunk-based RAG with structure-aware subgraph retrieval over document graphs, developed during a research internship at the Computer Vision Center (UAB, Barcelona) and published at ICDAR 2026.
NoTeS-Bank — a benchmark evaluating vision-language models on unstructured academic and scientific handwritten notes through evidence-based and open-domain VQA, published at ECML PKDD 2026.
Transformer-based handwriting recognition — a system that jointly leverages online (stroke) and offline (image) features, published at ACPR 2025.

I have collaborated with the Computer Vision and Pattern Recognition Unit (CVPRU) at the Indian Statistical Institute, Kolkata, working with Prof. Umapada Pal’s group on handwriting recognition and document analysis, and contributed to the ICDAR 2025 Handwritten Notes Understanding Challenge.

News

May 25, 2026	🎉 NoTeS-Bank has been accepted at ECML PKDD 2026! [paper] [arXiv]
May 15, 2026	🎉 My first-author paper From Chunks to Graphs: Training-Free Multimodal Late Interaction for Document Understanding has been accepted at ICDAR 2026 (Vienna, Austria)! [paper]
Dec 01, 2025	🌍 Started a remote research internship at the Computer Vision Center (UAB, Barcelona), working on multimodal RAG over document images with Dr. Sanket Biswas and Dr. Josep Lladós.
Nov 10, 2025	📄 My first-author paper A Transformer Based Handwriting Recognition System Jointly Using Online and Offline Features is out in the ACPR 2025 proceedings!
Sep 15, 2025	🏆 Our Handwritten Notes Understanding Challenge paper was presented at ICDAR 2025.

Research at a Glance

Publication Timeline

{
  "tooltip": { "trigger": "axis" },
  "grid": { "left": "5%", "right": "5%", "bottom": "10%", "containLabel": true },
  "xAxis": {
    "type": "category",
    "data": ["2025", "2026"],
    "axisLabel": { "color": "#666" }
  },
  "yAxis": {
    "type": "value",
    "name": "Papers",
    "minInterval": 1,
    "axisLabel": { "color": "#666" }
  },
  "series": [
    {
      "name": "Publications",
      "type": "bar",
      "barMaxWidth": 40,
      "data": [2, 2],
      "itemStyle": {
        "color": {
          "type": "linear",
          "x": 0, "y": 0, "x2": 0, "y2": 1,
          "colorStops": [
            { "offset": 0, "color": "#b93a14" },
            { "offset": 1, "color": "#e8a33d" }
          ]
        },
        "borderRadius": [4, 4, 0, 0]
      },
      "label": { "show": true, "position": "top" }
    }
  ]
}

Research Skills

{
  "tooltip": {},
  "radar": {
    "indicator": [
      { "name": "Computer\nVision", "max": 10 },
      { "name": "Deep\nLearning", "max": 10 },
      { "name": "Handwriting\nRecognition", "max": 10 },
      { "name": "Document\nUnderstanding", "max": 10 },
      { "name": "Multimodal\nVQA", "max": 10 }
    ],
    "radius": "52%",
    "axisNameGap": 8,
    "axisName": { "fontSize": 11, "lineHeight": 13 }
  },
  "series": [
    {
      "type": "radar",
      "data": [
        {
          "value": [8, 8, 9, 8, 7],
          "name": "Expertise",
          "areaStyle": { "opacity": 0.3 },
          "lineStyle": { "color": "#b93a14", "width": 2 },
          "itemStyle": { "color": "#b93a14" }
        }
      ]
    }
  ]
}

Publication Venues

{
  "tooltip": { "trigger": "item", "formatter": "{b}: {c} ({d}%)" },
  "legend": {
    "orient": "vertical",
    "right": "5%",
    "top": "center"
  },
  "series": [
    {
      "type": "pie",
      "radius": ["35%", "60%"],
      "center": ["38%", "50%"],
      "avoidLabelOverlap": true,
      "itemStyle": { "borderRadius": 6, "borderColor": "#fff", "borderWidth": 2 },
      "label": { "show": false },
      "emphasis": {
        "label": { "show": true, "fontSize": 13, "fontWeight": "bold" }
      },
      "data": [
        { "value": 2, "name": "ICDAR (LNCS)", "itemStyle": { "color": "#3b3486" } },
        { "value": 1, "name": "ECML PKDD (LNCS)", "itemStyle": { "color": "#e8a33d" } },
        { "value": 1, "name": "ACPR (LNCS)", "itemStyle": { "color": "#b93a14" } }
      ]
    }
  ]
}

Selected Publications

From Chunks to Graphs: Training-Free Multimodal Late Interaction for Document Understanding

Ayush Lodh, Souparni Mazumder, Sanket Biswas, Josep Lladós, and Nisha Singh Chauhan

In International Conference on Document Analysis and Recognition (ICDAR), Lecture Notes in Computer Science, 2026

Abs DOI Bib HTML

A training-free multimodal retrieval framework for document understanding that replaces chunk-based selection with structure-aware subgraph retrieval over Doc2Graph representations, improving factual consistency on document VQA over text-RAG baselines.
@inproceedings{lodh2026chunks, google_scholar_id = {9yKSN-GCB0IC}, title = {From Chunks to Graphs: Training-Free Multimodal Late Interaction for Document Understanding}, author = {Lodh, Ayush and Mazumder, Souparni and Biswas, Sanket and Llad{\'o}s, Josep and Chauhan, Nisha Singh}, booktitle = {International Conference on Document Analysis and Recognition (ICDAR), Lecture Notes in Computer Science}, pages = {592--609}, year = {2026}, doi = {10.1007/978-3-032-36039-7_35}, }
NoTeS-Bank: Benchmarking Vision-Language Models on Unstructured Academic and Scientific Note-Taking

Alloy Das, Sanket Biswas, Soumitri Chattopadhyay, Ayush Lodh, Aniket Pal, Priyanka Banerjee, Nisha Singh, C. V. Jawahar, Josep Lladós, and Dimosthenis Karatzas

In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD), Lecture Notes in Computer Science, 2026

Abs DOI arXiv Bib HTML

NoTeS-Bank is a benchmark for evaluating vision-language models on handwritten academic and scientific notes containing equations and diagrams. It defines two tasks — evidence-based VQA with localization and open-domain VQA with domain classification — to assess vision-language fusion, retrieval, and multimodal reasoning.
@inproceedings{das2026notesbank, google_scholar_id = {2osOgNQ5qMEC}, title = {NoTeS-Bank: Benchmarking Vision-Language Models on Unstructured Academic and Scientific Note-Taking}, author = {Das, Alloy and Biswas, Sanket and Chattopadhyay, Soumitri and Lodh, Ayush and Pal, Aniket and Banerjee, Priyanka and Singh, Nisha and Jawahar, C. V. and Llad{\'o}s, Josep and Karatzas, Dimosthenis}, booktitle = {Joint European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD), Lecture Notes in Computer Science}, pages = {700--718}, year = {2026}, doi = {10.1007/978-3-032-37676-3_40}, arxiv = {2504.09249} }

A Transformer Based Handwriting Recognition System Jointly Using Online and Offline Features

Ayush Lodh, Ritabrata Chakraborty, Palaiahnakote Shivakumara, and Umapada Pal

In Asian Conference on Pattern Recognition (ACPR), Lecture Notes in Computer Science, 2025

DOI arXiv Bib HTML

@inproceedings{lodh2025transformer,
  google_scholar_id = {u-x6o8ySG0sC},
  title = {A Transformer Based Handwriting Recognition System Jointly Using Online and Offline Features},
  author = {Lodh, Ayush and Chakraborty, Ritabrata and Shivakumara, Palaiahnakote and Pal, Umapada},
  booktitle = {Asian Conference on Pattern Recognition (ACPR), Lecture Notes in Computer Science},
  pages = {250--264},
  year = {2025},
  doi = {10.1007/978-981-95-4395-3_18},
  arxiv = {2506.20255}
}

ICDAR 2025 Handwritten Notes Understanding Challenge

Aniket Pal, Sanket Biswas, Alloy Das, Ayush Lodh, Priyanka Banerjee, Soumitri Chattopadhyay, Ajoy Mondal, Dimosthenis Karatzas, Josep Lladós, and C. V. Jawahar

In International Conference on Document Analysis and Recognition (ICDAR), Lecture Notes in Computer Science, 2025

DOI Bib HTML

@inproceedings{pal2025icdar,
  google_scholar_id = {d1gkVwhDpl0C},
  title = {ICDAR 2025 Handwritten Notes Understanding Challenge},
  author = {Pal, Aniket and Biswas, Sanket and Das, Alloy and Lodh, Ayush and Banerjee, Priyanka and Chattopadhyay, Soumitri and Mondal, Ajoy and Karatzas, Dimosthenis and Llad{\'o}s, Josep and Jawahar, C. V.},
  booktitle = {International Conference on Document Analysis and Recognition (ICDAR), Lecture Notes in Computer Science},
  pages = {553--567},
  year = {2025},
  doi = {10.1007/978-3-032-04630-7_32},
}