Evaluation
Result
In our experiment, we use three different weights and cfg file to process the detection. At first, we used the pretrain yolo version 3 weights from YOLO’s official website to detect the sample. Then we go to pretrain yolov3-tiny and yolov3-tiny trained by Udacity dataset to get better processing time and better accuracy.
YOLOV3 (pretrained)
YOLOV3-tiny (pretrained)
YOLOV3-tiny(trained from Udacity dataset)
Analysis
The following table shows the precision, recall and single image processing time for three different models list above:
Video Detection
With YOLOV3-tiny(trained from Udacity dataset), we can detect a image in a fair process time and an acceptable accuracy. We apply our implementation in video stream by process every frame within the video.
Video demo 1:
Video demo 2: