Training model, object detetion

domenico.depascale · September 7, 2023, 2:00pm

Hi,

Anyone can help me with a training issue?
I’m trying to get my own custom object detection on Jetson AGX Orin. I start from training the model with about 2k of images, performing about 300 epochs. The objects I want to detect are about 50. At the end of the process the most part of them are well recognized (very high %), but 1-2 are never detected, even if the number of images are the same for all objects.

What can be the problem?

AastaLLL · September 8, 2023, 3:03am

Hi,

We have several tutorials about training an object detector.
For example:

github.com

dusty-nv/jetson-inference/blob/master/docs/pytorch-ssd.md

<img src="https://github.com/dusty-nv/jetson-inference/raw/master/docs/images/deep-vision-header.jpg" width="100%">
<p align="right"><sup><a href="pytorch-collect.md">Back</a> | <a href="pytorch-collect-detection.md">Next</a> | </sup><a href="../README.md#hello-ai-world"><sup>Contents</sup></a>
<br/>
<sup>Transfer Learning - Object Detection</sup></s></p>

# Re-training SSD-Mobilenet

Next, we'll train our own SSD-Mobilenet object detection model using PyTorch and the [Open Images](https://storage.googleapis.com/openimages/web/visualizer/index.html?set=train&type=detection&c=%2Fm%2F06l9r) dataset.  SSD-Mobilenet is a popular network architecture for realtime object detection on mobile and embedded devices that combines the [SSD-300](https://arxiv.org/abs/1512.02325) Single-Shot MultiBox Detector with a [Mobilenet](https://arxiv.org/abs/1704.04861) backbone.  

<a href="https://arxiv.org/abs/1512.02325"><img src="https://github.com/dusty-nv/jetson-inference/raw/master/docs/images/pytorch-ssd-mobilenet.jpg"></a>

In the example below, we'll train a custom detection model that locates 8 different varieties of fruit, although you are welcome to pick from any of the [600 classes](https://github.com/dusty-nv/pytorch-ssd/blob/master/open_images_classes.txt) in the Open Images dataset to train your model on.  You can visually browse the dataset [here](https://storage.googleapis.com/openimages/web/visualizer/index.html?set=train&type=detection).

<img src="https://github.com/dusty-nv/jetson-inference/raw/master/docs/images/pytorch-fruit.jpg">

To get started, first make sure that you have [JetPack 4.4](https://developer.nvidia.com/embedded/jetpack) (or newer) and [PyTorch installed](pytorch-transfer-learning.md#installing-pytorch) for **Python 3** on your Jetson.  JetPack 4.4 includes TensorRT 7.1, which is the minimum TensorRT version that supports loading SSD-Mobilenet via ONNX.  Newer versions of TensorRT are fine too.

## Setup

The PyTorch code for training SSD-Mobilenet is found in the repo under [`jetson-inference/python/training/detection/ssd`](https://github.com/dusty-nv/pytorch-ssd).  If you aren't [Running the Docker Container](aux-docker.md), there are a couple steps required before using it:

This file has been truncated. show original

Please check if you can get the expected custom model by following our tutorial.
Thanks.

domenico.depascale · September 8, 2023, 6:19am

Dear @AastaLLL thank you for the answer.

Of course I know this tutorial and I followed it step by step from the beginning.
However I got the issue I talked before.

Please, Have you other suggestions about it?

Thank you very much

AastaLLL · September 18, 2023, 9:26am

Hi,

Does every category have a similar amount of training data?
Thanks.