Some points to use own data & converted darknet weights

[kazuya-n]

It's no longer issue, but we have to be careful with these points. To detect original dataset & use original trained darknet weights. I close this issue soon.

• [ ] convert.py works well on Python 2
• [ ] Specify collect size of trained (detected) image. https://github.com/Qidian213/deep_sort_yolov3/blob/c0a82cab0d0777432a9f1cd43ef0ba0d703f836d/yolo.py#L31
• [ ] Be careful that original code ignores everything but person. (We can simply delete these lines to detect all classes, but SORT algorithm is aimed to detect person.) https://github.com/Qidian213/deep_sort_yolov3/blob/c0a82cab0d0777432a9f1cd43ef0ba0d703f836d/yolo.py#L100-L101
• [ ] We can detect our own video data by changing this line like cv2.VideoCapture("detect.avi") https://github.com/Qidian213/deep_sort_yolov3/blob/c0a82cab0d0777432a9f1cd43ef0ba0d703f836d/demo.py#L39
• [ ] Origninal darknet yolov3's colorspace is RGB, but opencv's one is BGR, We must convert colorspaces to use trained data by darknet. https://github.com/Qidian213/deep_sort_yolov3/blob/c0a82cab0d0777432a9f1cd43ef0ba0d703f836d/demo.py#L57
  • We can convert colorspaces by changing this line like image = Image.fromarray(frame[:,:,::-1])

[Qidian213]

Thanks