ML_object_detection

Maintained by the Danish Climate Data Agency for Bounding Box Detection on Oblique Images.

Installation

• clone repo

  git clone https://github.com/SDFIdk/ML_object_detection

• create conda environment

  cd ML_object_detection
  mamba env create -f environment.yml
  mamba activate  ML_object_detection

Example use

• Use the included model for inference on one small and one large image from the included set of example images

  python src/ML_object_detection/infer_with_sahi.py --weights models/example_model.pt --folder_with_images data/example_images/ --result_folder output
  
  by placing the output .json files in the same folder as the image files you can inspect the result with labelme

Download dataset from Hugging Face and train

The KDS objects in oblique images dataset is available on Hugging Face in YOLO format (images, labels, and dataset.yaml). To download it and train:

1. Install the Hub client (if not already installed):

   pip install huggingface_hub

2. Download the dataset (e.g. into a folder named KDS_objects):

   python -c "
   from huggingface_hub import snapshot_download
   snapshot_download(
       repo_id=\"rasmuspjohansson/KDS_objects_in_oblique_images\",
       repo_type=\"dataset\",
       local_dir=\"./KDS_objects\"
   )
   "

   Or clone the repo: git clone https://huggingface.co/datasets/rasmuspjohansson/KDS_objects_in_oblique_images KDS_objects

3. Set the dataset path in dataset.yaml
   Open KDS_objects/dataset.yaml and set path to the directory that contains images/ and labels/ (the folder where you downloaded the data). For example, if you downloaded into ./KDS_objects:

   path: .    # use "." if you run train from inside KDS_objects, or use the full path to KDS_objects
   train: images/train
   val: images/val
   test:
   names:
     0: Velux
     1: Kvist
     2: Altan

   If you run training from the project root, set path to the absolute or relative path to KDS_objects, e.g. path: ./KDS_objects.

4. Train with the downloaded dataset:

   python src/ML_object_detection/train.py --data ./KDS_objects/dataset.yaml

   Add other options as needed (e.g. --epochs, --imgsz, --weights).

Download labelme format and convert to YOLO for training

The same dataset is also provided in labelme format in the labelme_format/ folder on the Hub (one .json and one .tif per image). To download that folder and convert it to YOLO so you can train:

1. Download the dataset (or clone the repo) so you have labelme_format/ locally, e.g.:

   python -c "
   from huggingface_hub import snapshot_download
   snapshot_download(
       repo_id=\"rasmuspjohansson/KDS_objects_in_oblique_images\",
       repo_type=\"dataset\",
       local_dir=\"./KDS_objects\"
   )
   "

   Then your labelme files are in KDS_objects/labelme_format/ (.json and .tif pairs).

2. (Optional) Standardize JSON if your labelme JSONs refer to different image paths or you need a consistent format:

   python src/ML_object_detection/standardize_json.py --json_dir ./KDS_objects/labelme_format

3. Convert labelme to YOLO with labelme2yolo (creates images/, labels/, and a config YAML):

   labelme2yolo --json_dir ./KDS_objects/labelme_format --val_size 0.15 --test_size 0.15

   By default the YOLO dataset is written under a subfolder of the directory that contains the JSON dir (see labelme2yolo --help for --out_dir if you want a specific path).

4. Set the dataset path in the generated dataset.yaml (e.g. path: . or the path to the folder that contains images/ and labels/).

5. Train using the generated YOLO config:

   python src/ML_object_detection/train.py --data ./KDS_objects/YOLODataset/dataset.yaml

   (Adjust the path to dataset.yaml to match where labelme2yolo wrote it.)

Create new dataset for training

• split the images to sizes suitable for yolo

  python split_with_gdal.py --image /path/to/large/images --output dataset/folder --x 640 --y 640 --overlap 40

• Create a dataset with labelme

  Open folder containing the splitted images

  For each image you want to train on, draw rectangles from the upper left corner to the lower right corner.

  OBS. All objects of the categoriez you want to detect needs to be marked up. Partly labeled images will ruin the training.

• (optional) set all "unkown"/"ignore" areas to black

  If you labeled areas with the text "ignore" (e.g areas for wich you are unsure about the correct classification) we have the option to mask all these areas and make them black. Calling mask_unknown_regions.py with -h flag will give more instructions on usage

• copy all data to a new location before doing the next steps

  note: draw rectangles from the upper left corner to the lower right corner

• (optional) set all "unknown"/"ignore" areas to black:

  python src/ML_object_detection/mask_unknown_regions.py -h

• make sure that all .json files use the same format (original .tif image)

python standardize_json.py --json_dir /mnt/T/mnt/trainingdata/object_detection/from_Fdrev_ampol/all/

• convert the labelme dataset to yolo format with

  labelme2yolo --json_dir /path/to/labelme_json_dir/ --val_size 0.15 --test_size 0.15
  e.g labelme2yolo --json_dir /mnt/T/mnt/trainingdata/object_detection/from_Fdrev_ampol/split/

• Train a object detection model on the dataset

  python train.py --data /path/to/labelme_json_dir/config.yml
  e.g  python src/ML_object_detection/train.py --data /mnt/T/mnt/trainingdata/object_detection/object_detection_dataset/2025-06-16/labelme_images/YOLODataset/dataset.yaml

• Use the model for inference on large (unsplitted images) (e.g for creating sugestions for new labels)

  python src/ML_object_detection/infer_with_sahi.py --weights models/example_model.pt --folder_with_images data/example_images/ --result_folder output