scripts/frames_to_MOT.py

import argparse
import torch
import os
import json
from tqdm import tqdm

import project_subpath
from backend.dataloader import create_dataloader_frames_only
from backend.inference import setup_model, do_detection, do_suppression
from backend.InferenceConfig import InferenceConfig

from lib.yolov5.utils.general import clip_boxes, scale_boxes


def main(args, config=InferenceConfig(), verbose=False):
    """
    Construct and save MOT format detections from yolov5 based on a frame directory
    Args:
        frames (str): path to image directory
        output (str): where MOT detections will be stored
        weights (str): path to model weights
    """
    
    print("In task...")
    print("Cuda available in task?", torch.cuda.is_available())

    print("Config:", config.to_dict())

    model, device = setup_model(args.weights)

    in_loc_dir = os.path.join(args.frames, args.location)
    out_loc_dir = os.path.join(args.output, args.location)
    metadata_path = os.path.join(args.metadata, args.location + ".json")
    print(in_loc_dir)
    print(out_loc_dir)
    print(metadata_path)

    detect_location(in_loc_dir, out_loc_dir, metadata_path, config, model, device, verbose)


                
def detect_location(in_loc_dir, out_loc_dir, metadata_path, config, model, device, verbose):

    seq_list = os.listdir(in_loc_dir)

    with tqdm(total=len(seq_list), desc="...", ncols=0) as pbar:
        for seq in seq_list:


            image_meter = (-1, -1)
            with open(metadata_path, 'r') as f:
                json_object = json.loads(f.read())
                for sequence in json_object:
                    if sequence['clip_name'] == seq:
                        image_meter = (
                            sequence['x_meter_stop'] - sequence['x_meter_start'],
                            sequence['y_meter_stop'] - sequence['y_meter_start']
                        )

            pbar.update(1)
            if (seq.startswith(".")): continue
            pbar.set_description("Processing " + seq)

            in_seq_dir = os.path.join(in_loc_dir, seq)
            out_seq_dir = os.path.join(out_loc_dir, seq)
            os.makedirs(out_seq_dir, exist_ok=True)

            detect_seq(in_seq_dir, out_seq_dir, image_meter, config, model, device, verbose)

def detect_seq(in_seq_dir, out_seq_dir, image_meter, config, model, device, verbose):

    ann_list = []
    frame_list = detect(in_seq_dir, image_meter, config, model, device, verbose)
    for frame in frame_list:
        if frame is not None:
            for ann in frame:
                ann_list.append({
                    'image_id': ann[5],
                    'category_id': 0,
                    'bbox': [ann[0], ann[1], ann[2] - ann[0], ann[3] - ann[1]],
                    'score': ann[4]
                })
    result = json.dumps(ann_list)

    with open(os.path.join(out_seq_dir, 'pred.json'), 'w') as f:
        f.write(result)

def detect(in_dir, image_meter, config, model, device, verbose):
    
    #progress_log = lambda p, m: 0

    # create dataloader
    dataloader = create_dataloader_frames_only(in_dir)

    inference, image_shapes, width, height = do_detection(dataloader, model, device, verbose=verbose)

    outputs = do_suppression(inference, image_meter_width=image_meter[0], image_pixel_width=image_meter[1], conf_thres=config.conf_thresh, iou_thres=config.nms_iou, verbose=verbose)

    file_names = dataloader.files
    frame_list = []
    for batch_i, batch in enumerate(outputs):

        batch_shapes = image_shapes[batch_i]

        # Format results
        for si, pred in enumerate(batch):
            (image_shape, original_shape) = batch_shapes[si]

            # Clip boxes to image bounds and resize to input shape
            clip_boxes(pred, (height, width))
            boxes = pred[:, :4].clone()  # xyxy
            confs = pred[:, 4].clone().tolist()
            scale_boxes(image_shape, boxes, original_shape[0], original_shape[1])  # to original shape

            frame = [ [*bb, conf] for bb, conf in zip(boxes.tolist(), confs) ]
            
            file_name = file_names[batch_i*32 + si]
            for ann in frame: 
                ann.append(file_name)

            frame_list.append(frame)
            
    return frame_list

def argument_parser():
    parser = argparse.ArgumentParser()
    parser.add_argument("--frames", required=True, help="Path to frame directory. Required.")
    parser.add_argument("--metadata", required=True, help="Path to frame directory. Required.")
    parser.add_argument("--location", required=True, help="Name of location dir. Required.")
    parser.add_argument("--output", required=True, help="Path to output directory. Required.")
    parser.add_argument("--weights", default='models/v5m_896_300best.pt', help="Path to saved YOLOv5 weights. Default: ../models/v5m_896_300best.pt")
    return parser

if __name__ == "__main__":
    args = argument_parser().parse_args()
    main(args)