shouryade / parking-yolo-tiny

Repository used for CNN in the WCE Hackathon 2022 | Rochak is a Madlad

Model: Yolov4-tiny Training Code: YOLOv4_tiny_Darknet_Parking.ipynb For training use your own roboflow api key.

Dataset : UFPArk raw data https://nextcloud.lasseufpa.org/s/qxWsqNjLSgqqdHM

Processed Dataset: https://app.roboflow.com/tiny-yolo-v4-y89yh/cnn-hack/1

Implementation: Yolo-in-cv2-video.ipynb 1 cell for Photos Implementation 2 cell for Video Implementaion

Model Details: calculation mAP (mean average precision)=96 detections_count = 3434, unique_truth_count = 2758
class_id = 0, name = empty, ap = 81.96% (TP = 936, FP = 159) class_id = 1, name = parked, ap = 98.90% (TP = 1560, FP = 84) class_id = 2, name = unauthorized, ap = 88.54% (TP = 21, FP = 2)

for conf_thresh = 0.25, precision = 0.91, recall = 0.91, F1-score = 0.91 for conf_thresh = 0.25, TP = 2517, FP = 245, FN = 241, average IoU = 78.84 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall mean average precision ([email protected]) = 0.898009, or 89.80 %

Impementaion Details: FPS 24-30 While displaying output FPS 45-50 While processing only