PP4AV data manipulation

Process data:

• Split dataset into train/val/test: edit image path, label path, and save directory in utils/config/dataset_split_config.yaml and run

python dataset_split.py 

• Convert data to Coco format (output of anonymizer model saved .txt and need convert to .json): edit config in utils/config/format_transform_config.yaml and run:

python format_transform.py  

• Non-maximum suppression: edit config in sample_configconfig/nms_config.yaml and run:

python nms.py 

• Statistic number of boxes, AP, AR of json ground-truth and json prediction file
  • input: json coco true file, json coco pred file corresponding
  • output: csv statistic file (each image is a row with number boxes, AP, AR of each class)
  • Edit config in analysis_stats_config.yaml: true_annotation_path is list of path to json coco true file, pred_annotation_path is list pf path json coco pred file corresponding, save_path is list of path to save csv file, nms_threshold, true_conf_threshold is confidence score threhold of coco true file, pred_conf_threshold is confidence score threhold of coco pred file, size_threshold is area threshold box removed (if area is smaller than size_threshold, remove it)

  python –m utils.analysis_stats