Image classifier of items from the video game 'Rocket League'.
* The project is currently on hold until I have enough knowledge to proceed with the multi-labeling classification :) *
First job with Artificial Intelligence.
This is a project to try to identify the type of different Rocket League items via images. I created a dataset collected, modified and cleaned from rlgarage.
All the images were 220x220px and .png format.
The images were separated into different directories according to the type they corresponded to. The training and validation directories were also added with 1584 images belonging to training and 395 images belonging to validation.
I had to remove the "explosions" element type, since it would create a big problem for the model when it comes to differentiating between "boosts" and "explosions", since in the images they were practically impossible to classify even for a person with knowledge about the game.
This is my first project related to AI, with CNN and image classification and if possible I would like to improve the scope of this project to deploy the model in a real environment when it is finished and is able to classify several items in a single photo.
For example, given this image:
The model predicted this:
Saving 42a71338bd1551122241.webp to 42a71338bd1551122241.webp
1/1 [==============================] - 7s 7s/step
42a71338bd1551122241.webp
This image is in: [wheels]
Version 0.1.1 is out! This version includes the following:
RL_itemsdataset available to download and use inside the notebook.86% accuracy example modeladded for easier understanding of the project.- Image augmentation included to try to improve the model's accuracy.
- Allows users to build their own model from a prebuilt example.
Weak points:
86% accuracy example modelshould be modified or redone to improve it's accuracy.Model predictionsshould be easier to understand for the final user, instead of showing the results with numbers.
Future Improvements:
Save & Download your modelfeature should be added.Model predictionsshould be easier to understand for the final user, instead of showing the results with numbers.- Finish the 'Conclusions' section in the notebook.
Few changes added!
96% accuracy example modelin order to substitute the old example model.- Image augmentation was removed as the main objective of the project is to identify the type of the object from a specific type of image, allowing the model to improve it's accuracy by
10%. Save & Download your modelfeature was added. Now you can continue working with your own creation!Model predictionsare now shown as the type of object that was predicted instead of a bunch of numbers.
Weak points:
- The model struggle to differenciate sometimes images between
boosts & trails. - There is no
Table of Contentsin the notebook.
Future Improvements:
- Add the version of the project to the notebook.
- Maybe I could improve the explanations of each block inside the notebook for the reader to understand it better.
- This is was the first milestone in the project, now I can start studying the possibilities to expand the model or create a new one to classify various items in the same image.
- Finish the 'Conclusions' section in the notebook.
tensorflow and tensorflow.keras for the model training and the model structure.
matplotlib.pyplot, numpy and PIL.Image for showing the results of the training and some examples of the dataset.
gdown and google.colab to load/download the model into your PC once it has been saved.
DeepLearning.AI TensorFlow Developer Professional Certificate - DeepLearning.AI on Coursera🔗
Sparse Categorical Cross-Entropy vs Categorical Cross-Entropy - Felipe A. Moreno - Medium🔗
Save, serialize, and export models - Neel Kovelamudi, Francois Chollet - TensorFlow🔗