solo2307/sketchmaps-detection

Siamese YOLOv9e – Sketch Map Marking Detection (Final Checkpoints)

Overview

This repository provides final YOLO checkpoints for detecting hand-drawn markings on printed maps. The models use a Siamese (dual-stream) YOLOv9e design for change detection: a clean basemap and an annotated map are processed jointly to detect only newly added markings.

All final checkpoints were trained with:

1. synthetic pretraining on OSM-based data, followed by
2. fine-tuning on hand-drawn datasets (two basemap domains)

---

Available Models (Final)

Checkpoint file	Fine-tuning domain	Intended use
siamese_yolov9e_osm_hand.pt	Hand-drawn on OSM basemaps	Best for OSM/cartographic backgrounds
siamese_yolov9e_ewi_hand.pt	Hand-drawn on EWI basemaps	Best for satellite/imagery-style backgrounds

Notes:

• Both models share the same synthetic OSM pretraining stage.
• EWI-derived imagery used for fine-tuning cannot be redistributed; only the checkpoint is provided.

---

Input Format (6 channels)

The models expect a 6-channel input representing a paired image:

• channels [0,1,2] = clean basemap (RGB)
• channels [3,4,5] = annotated map (RGB)

In other words: RGB_clean + RGB_annotated concatenated.

If your data is stored as multi-band TIFF:

• ensure channel order is RGB (not BGR)
• ensure the first 3 bands correspond to the clean map

---

Dataset

Training and evaluation datasets are available at: https://huggingface.co/datasets/solo2307/osm-sketchmaps