Better accuracy for object detction for similar classes

[music1913]

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc)
x86, Ubuntu, RTX3090.
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc)
Detectnet_v2
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)
• Training spec file(If have, please share here)
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)

I’m using a very small custom dataset for retraining based on un-pruned detectnet-v2.

my target detection classes are:

• eletric bicycle
• people

each class has less than 400 samples in training dataset.

as I’m testing with a video source with the export model by my deepstream app (detect and then upload above 2 classes to a remote server), I noticed the detection accuracy for eletric bicycle and traditional bicycle is pretty low, the model almost recognize all traditional bicycle to eletric bicycle.

I want to ask, what is the suggestion to help the model to better distinguish these 2 similar classes of traditional bicycle and eletric bicycle, as I can guess:

1. adding more training samples for eletric bicycle.
   then the model can better recognize the expecting object.
2. adding one extra classe of traditional bicycle, as well as training samples for it.
   when run the the inference, I can ignore this class by code.

[Morganh]

Yes, the two items you mentioned are suggested. We’d better add more training images during training.