Can someone tell me why the otherwise good training results drop so much after adding some new night and blurry data, and also why the bike detection rate is so low?
status.json (30.5 KB)
The TrafficCamNet model was trained on RGB images in good lighting conditions. Therefore, images captured in dark lighting conditions may not provide good detection results.
Some re-training will be required on these classes to improve accuracy.
1 Like
Is the low detection precision for two-wheelers for the same reason?
(ps: re-training means the retrain process in taotoolkit notebook?)
Very thx! And I saw the related post of the forum and raised the class weight para, but there doesn’t seem to be any improvement in accuracy
No. It means it is needed to train with your own datasets.
For your case, you can train with night and blurry data with ngc hdf5 pretrained model.
Is the low detection precision for two-wheelers for the same reason?
The recognition accuracy of two-wheelers is always low…
You can run below experiment to check if accuracy improves.
- Prepare two-wheelers images. And make sure the labels are correct.
- Split 15% to validation dataset. Others are training dataset.
- Generate tfrecords for them.
- Train for only one class : two-wheelers
- Set ngc pretrained model (hdf5 format)
Very thx! I’d like to know which ngc hdf5 pretrained model is more suitable?
Use the one for detectnet_v2. See Open Images Pre-trained DetectNet_v2 — TAO Toolkit 3.22.02 documentation
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.