FIRST: Machine Learning Options 2024

Technical Programming
, , ,
1

Hi there, we’re looking for the best affordable machine learning option for note tracking in 2024. We have Raspberry Pi 4B and Jetson TX1 in hand and can currently not afford another coprocessor like Google Coral or OrangePi. We’re thinking of writing a custom object detector API with TensorFlow or training YOLOv8 using Roboflow. No matter which one we use, we have problems with both processors since RPi works very slow and gets us very low FPS and Jetson takes up a lot of space. What do you think is the best solution?

3

Yolov8n (the nano version) runs 20fps on my very old laptop with intel i5 cpu
USE JESTON!!!

  1. use a coprocessor that used Navida GPU, then you can use cuda, that speeds up the process rate a lot
  2. You can also refactor it in cpp and use tensorRT to speed it up even more
1 Like
4

Also strongly suggest using yolov8n, use labelme to lable the dataset, write some custom script to convert from labelme format to yolo format, and use a server to train the modle

1 Like
5

Dataset Colab also has pretrained YOLOv8n models to download.

1 Like
6

@Rocky_S Thanks a lot for the answers!! Also I did some casual research on the topic but could you also explain the methods you mentioned please? I may especially need tips on refactoring in cpp

1 Like
7

Yes, they seem to help. Thanks!

2 Likes
8

I was kinda learning from my mom’s friend’s lab, and I leaned to do some very basic yolo things from a student their.
They used cpp and tensorRT and cuda to get 60fps plus
I’m not really sure how to do it either lol. I only use pretrained .pt and trained it on my own dataset

1 Like
9

you can download pretrained yolo from ultralytics,
The tool I used to label custom dataset is labelme, small terminal app, i might made a vitrualenv for it, can’t remember exactly
Roboflow is also a nice tool, might cost a little money for advanced features, idk

10

Strongly suggest using linux coprocessor, allows you to ssh and use vim to change the code in the pits, very helpful
getting reliable power supply for the coprocessor might be a challenge, a mentor helped us with that last year