Convert FLIR formatted annotations to YOLO training format + Additional Directory tree formatting guide.

• Converts FLIR json annotation file provided in the dataset,index.json, to YOLO labels for each image.
• Creates a .txt file with all image names extracted from JSON file and appends your specified directory to thier names.
• Previews Images with labels drawn with class names while converting for the user to visually check data validity.

1. Download and extract FLIR dataset (https://www.flir.eu/oem/adas/adas-dataset-form/)

2. Paste flir2yolo.py script inside a FLIR dataset folder ex: "images_thermal_train".

3. Run the script with the directory of the dataset in the YOLO folder added as an argument

python3 flir2yolo.py "data/thermal/train_flir_images/"

4. Redo the previous steps for all datasets that you need to convert.

5. Copy and rename the generated files to YOLO directory.

• A folder called labels will be created where the script resides inside the FLIR dataset folder. Which contiains the label txt files named by the same image names provided by FLIR ex: video-2Af3dwvs6YPfwSSf6-frame-000000-imW24bapJsHpTahce.txt. Each .txt file will contain the YOLO formated bounding boxes and classes class bx by bw bh ex: 2 0.603906 0.488281 0.323437 0.171875

• A .txt file called all_image_names.txt which will contain all images in the dataset file with the inputed directory appended to thier names ex: data/thermal/train_flir_images/video-2Af3dwvs6YPfwSSf6-frame-000000-imW24bapJsHpTahce.txt

This might help you with how the data should be prepared for training.

I have also included my .yaml config file which you may use or edit according to your naming.

├── YOLO
    ├── data
        ├── thermal
        │ ├── images
        │ │   ├── trian_flir_images   #contains all train images
        │ │   ├── test_flir_images    #contains all test images
        │ │   └── val_flir_images     #contains all val images
        │ ├── labels
        │ │   ├── trian_flir_images   #contains all generated train txt labels
        │ │   ├── test_flir_images    #contains all generated test txt labels
        │ │   └── val_flir_images     #contains all generated val txt labels
        │ ├── trian_flir_images.txt   #this is the all_image_names.txt generated file renamed
        │ ├── test_flir_images.txt    #this is the all_image_names.txt generated file renamed
        │ └── val_flir_images.txt     #this is the all_image_names.txt generated file renamed
        └── thermal.yaml

• I use opencv to display the labeled images, you may comment lines 6,42,61,63,64 if you dont want to display images
• I only needed 11 classes in my work, thus you will find a line that filters and only converts these classes. If you need more classes, just add them to the labels variable in the script.

28 labels = ['person','bike','car','motor','bus','train','truck','light','dog','scooter','other vehicle']

• The classes are numbered according to thier position in the list.

• This script is better than other scripts since this converts the orginial FLIR generated index.json file. The other coco.json file contains wrong labeling which will result in your training loosing hope.
• After the script ends, it will print the ammount of images dicovered and labeled. You will notice that in some datasets there will be way less displayed count than the number of images available. This is due to FLIR not labeling these images and not including them in the JSON file. Dont panic..
• Please talk encouraging words and bring cookies to your GPU while training for optimal performance.ʕっ•ᴥ•ʔっ