You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Jan 4, 2024. It is now read-only.
Hello! I follow these instructions in dianjixz/v831_yolo and set both train_size and val_size parameters are 416*416 and to train in my computer, and then I convert my mode into onnx mode and ncnn mode, finally I upload them on the toolbox website to convert into awnn mode.
`
I am sure that my training and modes are correct, and I modify below these codes and replace my trained mode, I try my best to run many time on V831 with C++ code, however there are some ISP errors occurring and the screen is full green(I upload the yolov2_camera_error.log and you can check it), how do I sovle it?
However I follow the same setting in Python code, it can run normally with low frame rate. Does the V831 not support the input_size with 416*416 in C++ code? I am urgent to apply these into my projects recently, thank you for your help!