felipesulser / computational-intelligence-lab---ethz

Computational Intelligence Lab ETHZ 2017

Testing

• deep network datafiles are in drive called "datafiles_deep"
• deep network raw output is in drive called "result_raw_deep"

Road Segmentation

Data set augmentation

• Rotate each image by 90 deg
• Choose 7 more images with more diagonal roads and highways: [23,26,27,42,72,83,91]
• Rotate these by 180 and 270 deg, so at the end we have 214 input images
• Reshuffle the data set
• Before balancing the data set, shuffle both classes again

Preprocess

• Histogram equalization failed
• Subtracting mean patch failed

Feature Selection

• Patch size x-y means that x is the core of the patch and y is the added context
• Label is determinated only for the core part
• Patch size: 16-42 works well
• Even better: 16-64
• We can try to run PCA on the patches to reduce the dimension from (y,y,3) to (y,y)

CNN Architecture

Shallow version

• 4 conv-pool layers of depths: 16, 32, 32, 64 and filter sizes: 5, 3, 3, 3
• Max-pooling after each conv. layer with ksize = strides = 2
• 3 fully-connected layers of depths: 48, 16, 2
• Outputs softmax
• Score obtained: 0.88629

Deep version

• 4 conv-pool layers of depths: 64, 128, 256, 512 and filter sizes: 3, 3, 3, 3
• Max-pooling after each conv. layer with ksize = strides = 2
• 3 fully-connected layers of depths: 2048, 2048, 2
• Outputs softmax
• Score obtained: still training

Post-processing

• Convolution of size 9x9 to smooth
• Binarize the image with threshold 0.5
• remove_filtering_neighbors() with 7 neighboors
• Total Variance denoising (TV) was OK, but the convoltuion one was better
• RandomForest with a window of 7x7 patches or 5x5 to predict the center's patch color was OK