Object Detection With Tensorflow

On one cold winter day in July … that is in Australia … I felt urgent need to train AI model for detecting uncapped concrete reinforcement bars…

So I bought from local Bunnings Warehouse store some new shiny reo bars like these

and bar safety caps like these

Started with caps

Took a hundred photos of different compositions with these yellow reo bar caps, labelled in LabelImg, took some SSD model, then another EfficientNet did a bit of coding in Python- and model quickly picked up what I wanted of her.

initially model took almost all yellow cylindric objects as reo bar caps

And boundary boxes could be more precise…

Then bars came in

Took several other useful tools from the garage and took another hundred of photos like this, and model started seeing better…

Yes, definitely better.

Mobile App

Finally, I trained a little SSD AI model to run it on my mobile phone without huge latency, created simple Android app and here how it looked like on Samsung S8:

It was a very good experiment.

Useful links

• MMdetection is not supported anymore
• Workflow for Training Object Detector with Label Studio & MMDetection
• LabelImg: https://github.com/HumanSignal/labelImg
• TensorFlow: https://www.tensorflow.org/
• Python Cheatsheet
• Conda Cheatsheet
• venv Cheatsheet
• Bash Cheat Sheet
• Ollama cheatsheet
• Generating PDF in Python - Libraries and examples"