Some improvement ideas

[debugmenot]

Hi!

1. num_classes and input_size for model should be available as model init parameters, because there can be multiple custom trained models of the same type (ex Yolov*) in single project, with different input size and num_classes. It will prevent code duplication.

2. Add YOLOv6 model for RKNN, for now, its most performant NN model for object detection for Rockchip. But RKNN-optimized YOLOv6 model has different opt_head that significately improves inference time (look at: https://github.com/airockchip/YOLOv6/commit/aa12cf70ee1795d00538f78a8857efd0cf17f7e4

3. Add RTSP (live Video) codec, it's almost the same as Video, except it takes not a Path but rtsp url

[Alwaysssssss]

Thank you for your suggestions.

1. We will continue to refine the parameters to accommodate various scenarios. In single project of YOLO, opening up num_classes and input_size can enhance versatility.
2. We will further deploy YOLOv6 based on RKNN.
3. We will integrate RTSP for improved usability in the future.