Trafficcamnet add truck class detection with other class and false detection occur.issue to detect small vehicles.miss detection

deepstream-5.1
tenserRT-.7.1
cuda-10.2
0s -18.04 ubuntu

in above trafficcamnet model miss small size .which is so fare in view vehicles and need to add truck class n existing python nvanalytics code

regards
gaurang patel

I find the background for your another topic created in deepstream forum. Video analytics. Will take a look for the detailed info.

Please suggest new pige config for improvement in detection and model to improve results.accuracy is 20 percentage out of 100 of detection.

Regards

Is the model trained by yourself or downloaded from ngc?

Download from ngc on 22/5/2022

Can you try Fp32 as well? More, seems that your test dataset is quite different from the training dataset of the ngc model. See more info in TrafficCamNet | NVIDIA NGC.
You can try to test some different dataset to double check.

Yes in config file FP32 change ?

No good results.

As mentioned above, your dataset has different data distribution against the training dataset of the ngc pretrained model.
To narrow down, suggest you to change to test the video which is already available in deepstream sdk.

Also, you can trim your test image and resize it to 960 x544.

I need to less change and improve results.
Suggest me.

Any changes in config file

i Need to trim video to 960*554 and then give to input or its directly possible by change in pgie_config file?

For your test video as below,

You can remove the black areas. Then change to 960*544 and then give to input.

Also, usually it is needed to train your dataset with TAO toolkit. The ngc trafficcamnet model may not work on any specific dataset which has different data distribution.

Why I need to trim input image 1280960 to 960544 for train own dataset? I need to use existing model first then it’s not working. Then i train model with my own dataset.

Please provide how to retain existing trafficnet model. Stepwise.

Not to train. Just to run inference. You can just trim the black areas, and save something similar to
image