Re-detection training with nvidia jetson nano

Autonomous Machines Jetson & Embedded Systems Jetson Nano
#1

Hi all

I want to do a pedestrian detection project with Nvidia Jetson Nano. I downloaded the jetson-inference repo. I ran Pednet with detectnet-camera.py but performance is not enough. There are examples of retraining classification networks in the jetson-inference repos but I couldn’t find an example of retraining for detection. I can not run train.py in the jetson-inference/python/training/detection. Can you tell me where to find the datasets for pedestrians? and can you tell me how to retrain pednet with these datasets via jetson?

#2

Hi @muse, instead of using pednet model, perhaps you might want to try the SSD-Mobilenet or SSD-Inception models that were trained on MS-COCO dataset. They are faster than pednet and include a class for people.

I am currently working on tutorial for object detection re-training, here is the PyTorch code awhile for it: https://github.com/dusty-nv/pytorch-ssd

There are also some links to TensorFlow training here: DIGITS or somthing else

#3

Thanks for your answer
I want the ssd-mobilenet-v2 and ssd-inception models to only detect human. I don’t want it to detect for other classes.
How can I do that?

#4

You would need to re-train it on only people. Here is part of pytorch-ssd repo that shows how to re-train: https://github.com/dusty-nv/pytorch-ssd#retrain-on-open-images-dataset

You can substitute whatever image classes you want that are available from the dataset. I am still working on the tutorial for it. For now, it may just be easier to ignore the classes you don’t want from the COCO-based models.

There’s also this pre-trained PeopleNet model available that you can run with DeepStream: https://ngc.nvidia.com/catalog/models/nvidia:tlt_peoplenet

#5

Thank you very much, but the repo is not enough. I still can’t re-train the ssd-mobilenet-v2 model. All I want is to make human detection. I want to delete other classes.

#6

You can follow that repo to re-train SSD-Mobilenet, and the SSD-Mobilenet-v1 is able to be exported to ONNX. You would get same/better performance with -v1 as -v2. As mentioned, I am also working on a more comprehensive Hello AI World tutorial around that repo. But for now, is there an error you are getting trying to run it?

#7

Hi
I want to run the pedestrian detection model below with jetson nano. Can i do that? İf yes, can you explain step by step?


Thank you
#8

Hi @muse, I haven’t tried caffe model of SSD-Mobilenet before, so I’m not sure if it would run with TensorRT or not (you could try trtexec from /usr/src/tensorrt/bin). Otherwise you would need to run it with the caffe fork linked to from that GitHub repo.