Sort-it

A machine-learning student project from 2020

Planet matters. Stream with your webcam or upload an image, and our application will tell you in wich trash you have to throw it.

Haar-Cascade Manually Trained

We created our own cascade classifier with Cascade Trainer GUI [1] on a dataset of 2527 positives and 6049 negatives. The training took place on a 4 cores i5-7300HQ @2.50GHz cpu chip running at full speed more than 4 hours. Best scale factor and close neighbours values seem to be around (1.5, 6). The .xml (cascade) can be found in the Data folder

Home-made Network Training

Thanks to Aadhav Vignesh [2], it has been possible to create and train our own convolutional neural network using pytorch. Running for hours on our CPU for 3 epochs, we obtained a model with an accuracy greater than 90%.

The .pt (learnable parameters of the network trained model) and the history.npy (plotting accuracy vs nb of epochs) can be found in the Data folder.

Instructions to use our WebApp

• Install requirements (see below)
• Download model_final_3_epochs.pt and cascadeGarbage.xml
• Download utils, detection, templates and convolution_neural_network
• Download main.py
• Launch main.py
• Go into your favorite browser and go the url localhost:5000
• Enter the correct settings in the project tab and use your favorite method

You can check our demonstration in our video.

Requirements

Please read requirements.txt and install necessary modules

To install pytorch, enter the following command : pip install torch==1.4.0+cpu torchvision==0.2.2 -f https://download.pytorch.org/whl/torch_stable.html

References

Images sources for haar cascade training :

• https://github.com/JoakimSoderberg/haarcascade-negatives

• https://github.com/garythung/trashnet

• Flickr API

Haar-Cascade classifier :

• [1] https://amin-ahmadi.com/cascade-trainer-gui/

• https://medium.com/@vipulgote4/guide-to-make-custom-haar-cascade-xml-file-for-object-detection-with-opencv-6932e22c3f0e

Deep Learning Network :

• [2] https://www.kaggle.com/aadhavvignesh/pytorch-garbage-classification-95-accuracy