Baby Steps of Object Detection

Computer vision has been a hot topic in the past decade and this as well, with machines being able to identify objects as humans, with some errors as well. Since the pandemic has invaded the world, the world has chosen to stay still with minimal interactions to avoid the spread and infections. This has in fact affected us a lot, being a hardware startup. We had our product ready to be shipped when this hit, and with all revenues dried, combined with the lack of supply of components, we thought of stepping into the world of machine learning which required minimal hardware components.

Google Inc has come up with TensorFlow, a platform for neural network training. We, have initially used ImageAI on Python for our initial testing. We coded, and then found that the codes were running fine and the machine was learning after a few failed attempts. The most difficult part, we though, would be to find a proper hardware to test the models and to train them.

We opted to create a library of the things we wanted the model to identify and started with it. We tried training the model from scratch. But then, we later learnt that there are pre trained models which we could use and then train them for our data set. This turned out to be faster.

We came across Google CoLab while we had to run the models to train for hours. Google CoLab is a free tool which can run these training with specialized GPU and TPUs. We mounted our Google Drive for data set and to save the trained models.

This is our journey so far. Stay tuned, as we move ahead.

The tools that we have used are

A snip from our training

Leave a Reply