In this experiment, Yolov8 nano and Yolov5 nano models were trained for car body parts detection.
Yolov8 results were higher than Yolov5 in nearly all aspects as inference time, mean average precision(mAP), and True Positives Rate (TPR);
however, Yolov8 took 6x the training time of Yolov5 on the same data, which reflects the extensive computational capabilities required for v8.
Train Time:
Both models were trained for 100 epochs, 16 batch size, 640 image size, 1840 train images, and 220 validation images.
- Yolov5: 30 Minutes, 36 Seconds
- Yolov8: 3 Hours, 14 Minutes, 42 Seconds
Inference Time:
At image shape (1, 3, 640, 640)
- Yolov5: 0.6ms pre-process, 11.2ms inference, 1.2ms NMS per image
- Yolov8: 0.3ms pre-process, 2.5ms inference, 0.0ms loss, 3.7ms post-process per image
Validation Confusion Matrix:
-
Yolov5:
-
Yolov8:
Validation PR Curve:
- Yolov5:
- Yolov8:
