What is a proper way to increase the accuracy of custom detectnet?

[AK51]

Dear Dusty,

I have create my custom detectnet, it works ok, thx. But I want to know the proper way to improve the accuracy. I use the same 1-minute.mp4 for training and to run the detectnet for checking.

Something I don't understand is that if there is a frame that the object cannot be recognized, I wrote down the time. I then video-capture that object at that time and put in the dataset for re-training. And I did delete the old folder from model folder in order to have a fresh start. (To be sure the engine file is gone.)

But when I run the detectnet again, that object still cannot be recognized. Is there any suggestion to recognize that object at that time? Note: I did carefully do the camera-capture and there is no wrong labelling and wrong box or empty box. (That is only one label) Note: The Loss is around 2.6 with 30 epoch Note: when I use the video-capture, I use merge set option (merge train, val and test) Thx