TrafficCamNet Model Retraining issue


I have retrained the TrafficCamNet pretrained model with our own custom data set which is having 5 classes ( car, bicycle, auto_rickshaw, truck, bus).

When we are using our retrained model with deep-stream application we observed that it is not detecting some class of objects which were able to detect with original NGC trafficcamnet model.

As you can see in the first picture all the objects are getting detected with original NGC TrafficCamNet model and our retrained model both because we have retrained the model with only front view of the objects

In the second image, you can see the objects are getting detected with original NGC TrafficCamNet Model but not getting detected with our retrained model. In our custom data set rear and side view of the objects is not there.

Now the question is why objects with side and rear view is not getting detected but earlier it was. Model should be cumulative i.e. It should have all the NGC TrafficCamNet features as well as our custom data set features too.

Please help us out to resolve this problem or correct me if I am having wrong understanding of transfer learning.

Thanks and Regards,
Vikas Dwivedi

How many times did you train? Can you share the spec file when you first time train with trafficcamnet model as the pretrained model?


Hi i am using NGC detectnet v2 Notebook for transfer learning on TrafficCamNet. I have used TrafficCamNet unpruned model for training.

i have done training one time. and then to get back accuracy i am retraining it according Notebook steps.

Bellow is My Spec file for training:
detectnet_v2_train_trafficcamnet.txt (6.5 KB)


Please try to evaluate your custom data with the tlt model you 1st time trained.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.