Real-Time Object Detection with TensorFlow and OpenCV

In the final project (ShoppingEye - New user experience) of my cohort (Tobi, Peter and me) I created an object detection with four new trained objects. 

For setting up and configuring our system, the following tutorials were very helpful:
- Gilbert Tanner - TensorFlow Object Detection
- Harrison Kinsley - Tensorflow Object Detection API
- Dat Tran - How to train your own Object Detector with TensorFlow’s Object Detector API

We used the following dependencies:

We decide on these four objects:

After I collected all images (approx. 150 images for each object), I had to prepare them accordingly (resize, split in train and test) and labeling them by hand. For these process I used LabelImg. Each label was saved in a separate XML-file and merged together into one CSV-file to convert them into a Record-file afterwards.

Now, after I created the required input file I was able to train my model. Here are some tests:

With the Tensorboard you can see how well your model performs, as you can see here in this screenshot. Our model learns up to 12k steps (epochs) very well and afterwards the result get worse. This is the point where our model overfits.

The project with the new trained model is available on my Github repo:

Posted in: