Custom network for DriveNet

Hi,

I have problems with my custom network and DriveNet.

I have trained a multiclass network using DIGITS and GoogleLeNet. while implementation, I notices dwDetector is for single class and DriveNet is for multiclass object detection but DriveNet uses NVIDIA proprietary DriveNet architecture.

How can I use my custom network for multiclass object detection using DriveNet or dwDetector?

Any help please?

Regards
Mayank

Dear mayank,
DriveNet APIs are to load and use our proprietary DriveNet model.
Firstly, to use any of your custom network with Driveworks API, you should get tensorRT model using TensorRT_optimization tool. We have added support for new layers in tensorRT and UFF model support in Driveworks in latest release. Could you please upgrade to latest PDK(https://developer.nvidia.com/nvidia-drive-downloads)

Dear mayank,
dwObjectDetector is actually generic and supports both single class and multi class. dwObjectDetector_getDetectedObjects API requires class index. so you can get detections for that class by providing the index. Please check process results in drivenet/common/DriveNet.cpp for usage.

Hi,
Thank you for the reply. I will keep you posted if it works.

Cheers
Mayank

Hi Siva,

I tried the dwObjectDetector_getDetectedObjects() on sample_object_dwdetector for 2 classes. I have put dwObjectDetector_getDetectedObjects() and drawBoxesWithLabels in a loop., but I am only able to get output for one class. There is no output for second class, but my model is trained on 2 classes using Digits.

Any idea what I am doing wrong?

Regards
Mayank

Dear mayank,
Is it possible to share code and model file here or via bug. I would like to reproduce it and check the issue.

Dear SivaRamaKrishna,

I have trained a 2 class Detectnet model using Digits and then have optimized it using the TensorRT optimization tool provided with Driveworks.

  1. Should I modify the sample_object_detector or modify sample_drivenet code to use the above multiclass network.
  2. Secondly what is the use of DW_DRIVENET_MODEL_CUSTOM in driveNet.h is it useful for inferencing other networks like Detectnet as trained above?