Always get 0 or nan precision during training detectnet

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc)
• Network Type (Detectnet_v2)
• TLT Version (docker_tag: v3.0-py3)
• Training spec file(SPECS_train.txt)
• How to reproduce the issue ? (tlt detectnet_v2 train -e /workspace/openalpr/SPECS_train.txt -r /workspace/openalpr/exp_unpruned -k nvidia_tlt)

I follow the steps on https://developer.nvidia.com/blog/creating-a-real-time-license-plate-detection-and-recognition-app to train my own data set, But Mean average_precision (in %): 0.0000 or nan, except that the data set is my own, I modify nothing ，I use my Iphone take the img.
lpd.tar.gz is my datasetslpd.tar.gz (4.1 MB)

tfrecord.log is tlt detectnet_v2 dataset_convert -d /workspace/openalpr/SPECS_tfrecord.txt -o /workspace/openalpr/lpd_tfrecord/lpd output log tfrecord.log (3.9 KB)

train.log is tlt detectnet_v2 train -e /workspace/openalpr/SPECS_train.txt -r /workspace/openalpr/exp_unpruned -k nvidia_tlt output logtrain.log (260.3 KB)

Hope someone help me,very thanks

3

It is related to the small lpd object. Most of the lpd objects are 10x10 pixels or less than it.
For detectnet_v2, it cannot detect this kind of pixels well. More info, see NVIDIA TAO Documentation

4

8/5000

@Morganh Thank you very much for your reply,I’m trying to detect and recognize license plates from a far distance,could you give me some advice.

5

Suggest enlarging the pixels of lpd objects.
Or you can try Yolo_v4 network.

6

Thank you very much,I will have a try.