You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

AGDSA Table Detection (YOLO, fine-tuned)

A YOLO object-detection model fine-tuned to detect table regions in scanned/rendered pages of financial report PDFs (customer-specific fine-tune, not a general-purpose table detector).

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict(image, imgsz=640, conf=0.3)

Detects a single class: table, as a bounding box in pixel coordinates.

Comparison vs. a general-purpose table-detection model

The examples below run this AGDSA-tuned YOLO detector (left) side by side with a general-purpose table-detection model (right, trained on PubTables-1M) on pages from public Australian government financial reports — documents outside the AGDSA customer data this model was fine-tuned on. Across these examples the fine-tuned detector matches or outperforms the general model on both detection recall and false-positive rate:

ASIC Annual Report 2024–25, page 6 — a real table is present; the AGDSA-tuned model detects it, the general model misses it (false negative).
Office of Hydrogen Power SA, financial statements, page 19 — a real table is present; the AGDSA-tuned model detects it, the general model misses it (false negative).
Office of Hydrogen Power SA, financial statements, page 16 — this page has no table (plain text policy notes); the AGDSA-tuned model correctly reports none, but the general model incorrectly detects one (false positive).

Downloads last month: -