Back to Projects

ADOPD

Ongoing work on understanding and generating multimodal documents

ADOPD is an ongoing research line on document intelligence. It begins with large-scale page decomposition and extends those structured page elements into grounded evidence for multimodal reasoning.

2024

A Large-Scale Document Page Decomposition Dataset

The original ADOPD introduced 120,000 visually diverse document pages spanning more than 1,000 document types, with dense annotations for document structure and semantic page elements.

Project page · Dataset

ADOPD: A Large-Scale Document Page Decomposition Dataset (ICLR 2024 poster).

2026

Thinking with Anchors: Grounded and Efficient Document Reasoning

ADOPD 2026 extends the original page decomposition dataset with human-cleaned captions, fine-grained semantic tags, and reasoning traces grounded in visual anchors such as text boxes and polygon masks. The goal is to move document understanding beyond localization toward evidence-backed multimodal reasoning.

Paper · Hugging Face collection

Data and Model Zoo

  • ADOPD2026 contains 120,000 high-resolution document pages with supervision for text detection, semantic region grounding, instance segmentation, page description, and document-level analysis.
  • DocCount is a manually curated visual-counting benchmark with 442 document images and explicit semantic counting questions.
  • Released models and demos include nine checkpoints across SAM 3/3.1, YOLOv12, RF-DETR, and LocateAnything-3B for detection, segmentation, and grounded document outputs.

Demos

Parallel, streaming generation of document structure — bounding boxes and polygon masks over a full page.
Agentic detection with self-refinement: over-recalled regions are grouped and merged into grounded evidence.

This work is currently under review; more details will be shared once it is published.

Back to Projects