DetectNetV_2: Object Detection

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

I have been using the Nvidia’s tlt:
DetectNet with googlnet

I am training for 5 classes and following are the instances of each classes in 10,770 train images:

  1. person [instance: 18,004] precision after training: 80.1
  2. gun [instance: 5056] precision after training: 52.92
  3. mask [instance: 4121] precision after training: 74.28
  4. face [instance: 9053] precision after training: 86.22
  5. helmet [instance: 2164] precision after training: 89.308
  6. knife [instance: 1206] precision after training: 48.36

Mean average_precision: 71.872%

But I could only get an overall precision of 68% initially, after which I tried to improve the annotations on images and could only get a mean average precision of 71.87% over all.
How do I resolve this issue. Please provide in-depth suggestions.

3

Hi @hitesh.pant21,
We recommend you to raise the concern in TLT forum to get better assistance.

Thanks!