Pets Detection And Segmentation Detectron2

Dataset

The Oxford-IIIT Pet Dataset: https://www.robots.ox.ac.uk/~vgg/data/pets/

Requirements

(Note: All the notebooks were executed on Google Colabs)

• Detectron2
• numpy==1.17
• pytorch
• opencv
• cocoapi==2.0

Prepare data

Run all blocks in prepare_data.ipynb to generate train_segmentation.json, test_segmentation.json, train_object_detection.json and test_object_detection.json

Train models

• Run the first 2 blocks in train.ipynb to install the required packages (if you use Colabs, you need to restart runtime after installation to update the environment).
• You need to specify the paths to your image directory, train_{segmentation/detection}.json, test_{segmentation/detection}.json and register the train and test set
• Pick a model config (.yaml file) in Detectron2's model zoo (use COCOInstanceSegmentation models for segmentation task and COCOObjectDetection for detection task) and set it via cfg.merge_from_file()
• Train the model

Evaluate models

• Run the first block in evaluate.ipynb to install the required packages (if you use Colabs, you need to restart runtime after installation to update the environment).
• Register dataset and set config
• Load trained weights using DefaultTrainer() class and resume checkpoint by setting trainer.resume_or_load(True)
• Execute function inferrence_on_dataset() to get the result
• Run the last block to visualize on images