DETR Doc Table Detection

detr-doc-table-detection is an object detection model that detects both bordered and borderless tables in document images. It is a fine-tuned version of facebook/detr-resnet-50, the DETR (Detection Transformer) architecture introduced in End-to-End Object Detection with Transformers. The model is hosted on gigarouter as a managed, OpenAI-compatible API.

Architecture and Training

DETR treats object detection as a direct set prediction problem, eliminating hand-designed components such as non-maximum suppression and anchor generation. The model uses a transformer encoder-decoder with a fixed set of learned object queries and a bipartite-matching loss. This fine-tuned model was trained on the ICDAR2019 Table Dataset, enabling it to recognize both bordered and borderless table structures in documents.

Performance Context

The underlying DETR architecture (ResNet-50 backbone) achieves accuracy and run-time performance on par with the highly-optimized Faster R-CNN baseline on the COCO object detection dataset, as reported in the original paper. While specific table-detection benchmark numbers for this fine-tuned variant are not published, the architectural strengths—global context reasoning and parallel prediction—are well suited to the diverse layouts found in documents.

About the Hosted API

No installation or model loading is required. Call the gigarouter endpoint with a document image and receive bounding-box detections for tables. The model handles both bordered and borderless tables, making it suitable for downstream tasks such as table extraction and document parsing.