Train Jetson-inference for cars,bikes,pedestrians,bus,motorbikes

Hi ,

I am using the detectnet code for camera on JetsonTX2 from Jetson-inference algorithm -
I want the detectnet code to detect for pedestrians,cars,bikes,bus and motorbikes.
The VOC-pascal-2005 dataset contains the above classes.
How to interface the Jetson-inference algorithm to get trained and detect objects for the VOC-pascal-2005 model?


Please prepare a desktop GPU for training use case first.

Here is our tutorial to help you set up the DIGITs environment to train a custom model:


Thanks for your response AastaLLL.
Can DIGITs be downloaded on NVIDIA Jetson TX2 to create a new model that can be used for live object detection(detectnet-camera)?


Hi Pratosha, DIGITS isn’t officially supported on Jetson. To keep training times manageable it’s standard to use a PC with discrete GPU or AWS/Azure in the cloud for DIGITS training. However someone from the community had reported getting DIGITS to run on Jetson TX1:

Not sure if it was reproducible or still works, perhaps you can comment if you decide to try it with JetPack 3.2.

Thanks for your response @dusty_nv.
Is there any other way I can train my own dataset to generate a model using JetsonTX2 that works on - (detectnet-camera).

Reason - JetsonTX2 is the only platform available for me to work on.

Kindly help.


It’s not recommended to use TX2 for training.
If other solution is not available for you, you still can set up nvCaffe for training directly.

Here is a tutorial for NvCaffe on Jetson:



Thankyou AastaLL,
I will definitely install and give it a try.
Just another query - Can desktop containing the GPU - GK107(GeForce GT 640 OEM) be used for installation of DIGITs 4?


Probably cloud is the way to go! I know that Nvidia has DIGITS on AWS marketplace.

You can even try and set up an instance on GCP with single GPU. The advantage? There is a $300 signup credit on GCP.

I used the latter option, still have some credits left! :)

It looks like the GeForce 640 OEM GPU card came in variants with 1GB or 2GB memory. Due to the small memory size, it would be preferred to use a cloud provider like bhargavK mentioned, as 1 or 2GB memory may not be enough for training some networks. To attempt to get training to fit in memory, you can try tweaking the Batch Size and Batch Accumulation hyperparameters like in this section of the tutorial: