Train a people recognition with Jetson Nano 4GB

usuario1918 · February 16, 2023, 7:43pm

Hello. I would like to train a model with about 5 photos of me so that the AI knows that it is me. I plan to retrain him every time he changes my clothes so he knows it’s me. Since there are 5 photos, I imagine that the process will be simple and with the people detector I can easily cut the photos.

Ref: Jetson AI Fundamentals - S3E5 - Training Object Detection Models - YouTube

Aim? I want you to identify me among several people.

Tell me what you think please. Do you see any error?

dusty_nv · February 16, 2023, 8:37pm

Hi @usuario1918, I think you will need more photos in your dataset. In my basic tests, I use a minimum of 100 images per class. And if you want it to be robust to different lighting conditions, camera angles, ect you may need to collect more than that.

usuario1918 · February 16, 2023, 8:51pm

I want to make a robot that will follow me when I ask. For that I look for him to only follow me with the clothes I currently have. That is to say, every time I tell it to follow me, it will take photos of me in seconds, it will train itself automatically and it will start to follow only me because it only recognizes me, that’s why I need them to be few, to train it on each request in a few seconds.

Don’t you think my idea can work (I’m installing everything to test)? Do you know any solution (maybe REID, I don’t know)?

usuario1918 · February 16, 2023, 8:56pm

I thought about using centroid tracking (GitHub - mailrocketsystems/Jetson-Nano-Person-Counter) but, it’s not that perfect, it crashes a lot, so I just want it to recognize me as the owner with a new trained model of me (just one class/label).

usuario1918 · February 16, 2023, 9:58pm

github.com

opencv/opencv/blob/master/samples/dnn/person_reid.py

#!/usr/bin/env python
'''
You can download a baseline ReID model and sample input from:
https://github.com/ReID-Team/ReID_extra_testdata

Authors of samples and Youtu ReID baseline:
        Xing Sun <winfredsun@tencent.com>
        Feng Zheng <zhengf@sustech.edu.cn>
        Xinyang Jiang <sevjiang@tencent.com>
        Fufu Yu <fufuyu@tencent.com>
        Enwei Zhang <miyozhang@tencent.com>

Copyright (C) 2020-2021, Tencent.
Copyright (C) 2020-2021, SUSTech.
'''
import argparse
import os.path
import numpy as np
import cv2 as cv

This file has been truncated. show original

https://drive.google.com/drive/folders/1wFGcuolSzX3_PqNKb4BAV3DNac7tYpc2

dusty_nv · February 16, 2023, 10:15pm

I haven’t tried that REID project, but yes it is good you experiment with some other methods like in OpenCV, one-shot learning, DeepSort tracking, ect given that you want it to learn/track almost instantly.

usuario1918 · February 16, 2023, 10:26pm

If you know of any technology that simplifies my work, I’ll appreciate it (I’m researching).

For example, when you mention this about one-shot learning, python’s face_recognition library came to mind. With a photo it can detect if it’s you or not, but, I need that for the whole body (or part of the body).

system · March 21, 2023, 7:57am

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