Detect damaged car parts using instance segmentation, mask r-cnn, and the Detectron2 library.
Using the kaggle dataset by LPLENKA, there are 59 images for training, 11 for validation, and 8 images for testing
- Kaggle: Go to COCO Car Damage Dataset by LPLENKA, create code with that dataset and import the kaggle notebook.
- Google Colab: Create your Kaggle API Token, upload this notebook to google colab, and dont forget to upload your API Token in the notebook.
Detectron2 is Facebook AI Research's next generation library that provides state-of-the-art detection and segmentation algorithms. It is the successor of Detectron and maskrcnn-benchmark. It supports a number of computer vision research projects and production applications in Facebook.
Context
The dataset contains car images with one or more damaged parts. The img/ folder has all 80 images in the dataset. There are three more folders train/, val/ and test/ for training, validation and testing purposes respectively.
Folders
train/:
Contains 59 images.
COCO_train_annos.json: Train annotation file for damages where damage is the one and only category. COCO_mul_train_annos.json: Train annotation file for parts having damages. There are five categories of parts based on which part the damage has happened. The parts can be namely, headlamp, front_bumper, hood, door, rear_bumper.
val/:
Contains 11 images.
COCO_val_annos.json: Validation annotation file for damages where damage is the one and only category. COCO_mul_val_annos.json: Validation annotation file for parts having damages. There are five categories of parts based on which part the damage has happened. The parts can be namely, headlamp, front_bumper, hood, door, rear_bumper.
test/:
Contains 8 images.
Annotation files have the following keys:
"annotations": Contains the bounding box and segmentation array. "categories": Contains the list of categories in the annotation. "images": Details of each image used in the annotation. "info": Creator information "licenses": License information