Case and members
Members: Ignacio Bengoechea
Kaufland data must be under subdirectoy Kaufland
Train and Test folder must be created empty
Preparation of the data
The train/test folders have been created with the sample core given by the mentor mitzev.
The train/test dataset images have been resized to 320×240 grayscale. This give us 76800 pixels.
This web because i couldn’t process al the images on my computer.
The dimensions of the data have been reduced using PCA to 1000 pixels. To show fast results.
I have used a simple neural network to classify the object.
This NN has 1000 inputs, the same number as the PCA output
This NN has 53 outputs, thats the number of clases of the trainf/folder
The value of epochs, learning rate and neurons have been selected using a hyperparameter tuning, not included in the final code.
Evaluation and results
Sorry, the accuracy is poor, a 2.78%.
That’s because this is a simple model, and we have over 53 categories in training dataset. The accuracy is almost random.
You can download the jupyter notebook here:
Thank you for your support. This is an amazing competition.
Due to the nature of the objects, I believe you will get better results if you keep some of colors in the dataset.
Good first attempt. I think you can improve on your Neural network architecture. Read more about convolutional neural networks they are a great help with image recognition tasks
Thank you very much for the feedback. I really apreciate it.
If you kept the colors you should get better results. It is excellent that you shared all you code in the notebook. You should also try to import additional datasets and enrich the provided data.
Good work! Simple and Elegant. You can slowly build on it with some other architecture like CNNs! Definitely good work to start with!