A submission for drivendata.com 3D-MRI video contest.
team: ru_kola
members: @mustafahakkoz, @Aysenuryilmazz
rank: 44 / 922
score (matthew's correlation coefficient): 0.1564
dataset: Stall Catchers micro dataset 2399 3D-MRI videos with 70/30 class ratio (flowing/stalled)
The main idea of the project is preprocessing videos hardly by blending all frames in to single one with different approaches, then applying a CNN classifier on them.
3 convolution+maxpooling dual layers followed by 1 fully-connected layer just as described in keras.io blog trained on kaggle's gpu kernels.
- Processing datasets and to images with 3 methods defined above: This code also produces predictions without ML, just by counting contours. process_videos.py
- CNN predictor: CNN.py
- Submission format for no-ML predictions: submission_format_no_ml.py
- Submission format for CNN predictions: submission_format_cnn_predictor.py
- Weights of trained CNN model: method1-nomedian-wholedataset.h5
- PCA debugging code for contour orientation method 1: pca_test.py
- PCA debugging code for contour orientation method 2: pca_test2.py
- Debugging code for whole preprocessing pipeline: readvideo2.py
- A sample video and output images: ./images and samples/