a segnet-like architecture for building detection in the spacenet dataset

spacenet_segnet.py

#from __future__ import absolute_import
from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img

from keras.callbacks import ModelCheckpoint
from keras.models import Sequential
from keras.layers import Convolution2D, MaxPooling2D
from keras.layers import Layer, Dense, Dropout, Activation, Flatten, Reshape, Merge, Permute
from keras.layers import ZeroPadding2D, UpSampling2D
from keras.layers.normalization import BatchNormalization
import sys
import os
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
#from preprocessing.visualize_prepro import shiftedColorMap
import itertools
import tensorflow as tf


path = sys.argv[1]
# input image dimensions
img_rows, img_cols = 400, 400
# output image dimensions
label_rows, label_cols = 400, 400

with tf.device('/gpu:1'):
    # we create two instances with the same arguments
    img_data_gen_args = dict(
        # featurewise_center=True,
        # featurewise_std_normalization=True,
        rescale=1. / 255,
        rotation_range=90.,
        width_shift_range=0.1,
        height_shift_range=0.1,
        zoom_range=0.2,
        fill_mode="constant",
        cval=0
    )
    label_data_gen_args = dict(
        rotation_range=90.,
        width_shift_range=0.1,
        height_shift_range=0.1,
        zoom_range=0.2,
        fill_mode="constant",
        cval=1
    )
    image_datagen = ImageDataGenerator(**img_data_gen_args)
    mask_datagen = ImageDataGenerator(**label_data_gen_args)

    # Provide the same seed and keyword arguments to the fit and flow methods
    seed = 1
    # image_datagen.fit(images, augment=True, seed=seed)
    # mask_datagen.fit(masks, augment=True, seed=seed)

    image_generator = image_datagen.flow_from_directory(
        os.path.join(path, '3band/'),
        target_size=(img_rows, img_cols),
        class_mode=None,
        batch_size=8,
        shuffle=False,
        seed=seed)

    mask_generator = mask_datagen.flow_from_directory(
        os.path.join(path, 'labels'),
        target_size=(label_rows, label_cols),
        class_mode=None,
        batch_size=8,
        shuffle=False,
        color_mode='grayscale',
        seed=seed)

    # combine generators into one which yields image and masks
    train_generator = itertools.izip(image_generator, mask_generator)

    kernel = 3
    filter_size = 64
    pad = 1
    pool_size = 2

    model = Sequential()

    model.add(Layer(input_shape=(img_rows, img_cols, 3)))
    # encoding layers
    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(filter_size, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D(pool_size=(pool_size, pool_size)))

    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(128, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(pool_size, pool_size)))

    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(256, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(pool_size, pool_size)))

    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(512, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))

    # decoding layers
    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(512, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())

    model.add(UpSampling2D(size=(pool_size, pool_size)))
    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(256, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())

    model.add(UpSampling2D(size=(pool_size, pool_size)))
    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(128, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())

    model.add(UpSampling2D(size=(pool_size, pool_size)))
    model.add(ZeroPadding2D(padding=(pad, pad)))
    model.add(Convolution2D(filter_size, kernel, kernel, border_mode='valid'))
    model.add(BatchNormalization())

    model.add(Convolution2D(1, 1, 1, border_mode='valid',))
    print model.output_shape
    model.add(Reshape((label_rows * label_cols,)))
    model.add(Activation('sigmoid'))

    model.add(Reshape((label_rows, label_cols, 1)))
    model.compile(loss="binary_crossentropy", optimizer='rmsprop',
                  metrics=['binary_accuracy'])
 
    model.summary()
    
    checkpointer = ModelCheckpoint(filepath="weights.hdf5", verbose=1, save_best_only=False)
    
    model.fit_generator(
        train_generator,
        samples_per_epoch=1000,
        nb_epoch=20,
        callbacks=[checkpointer])
    model.save('spacenetmodel2.h5')

Hi

I'm currently doing something similar and because I'm new to the subject of machine learning, I've got a question regarding the masks:

If we want to identify buildings, we've got typically a single image with a lot of instances (the buildings). If the network then requests the mask(s) for a specific image, do I have to return a single image which consists of all the masks/buildings on this single image or do I have to return an image where only one building is masked at a time?

Thanks for your help!

Hello,
I would like to use your model for road detection. However, I get the following error from the Reshape function in line 128:
ValueError: total size of new array must be unchanged
The output of the print(model.output_shape) statement is (None, 0, 0, 1).
Do you know what the problem might be?
Thank you in advance!

Hi

I'm currently doing something similar and because I'm new to the subject of machine learning, I've got a question regarding the masks:

If we want to identify buildings, we've got typically a single image with a lot of instances (the buildings). If the network then requests the mask(s) for a specific image, do I have to return a single image which consists of all the masks/buildings on this single image or do I have to return an image where only one building is masked at a time?

Thanks for your help!

• I think you should return a single image which consists all the masks/ buildings