+ str(name) + "') ..."
# Load large aerial view image and corresponding building and highway labels.
img = np.asarray(Image.open(GROUND_TRUTH_IMAGES_ROOT + name))
# np.unique(img[5471:6000, 5471:6000, :])
# for i in range(tiles_img.shape[0]):
# for j in range(tiles_img.shape[1]):
# print str(i * 12 + j + 1) + " unique: " + str(len(np.unique(tiles_img[i, j, :, :, :])))
# Get name of label file that corresponds to img.
label_file_name = name.replace('RGB.png', 'label_ETH_transformed.npy')
labels = np.load(GROUND_TRUTH_LABELS_ROOT + label_file_name)
# Split large aerival view image and np.ndarray of labels into tiles.
tiles_img = lib.image_to_tiles(img, lng_tiles, hgh_tiles)
tiles_labels = lib.image_to_tiles(labels, lng_tiles, hgh_tiles)
# Determine shape of the split of 'img' into tiles.
dim_tiles = lib.tiles_in_image(img, lng_tiles, hgh_tiles)
for i in range(0, dim_tiles[0]):
for j in range(0, dim_tiles[1]):
# Calculate the identifier for this tile.
identifier = str(number_tile_out).rjust(nr_of_digits_identifier, '0')
# Save tile image as numpy.ndarray.
np.save(TILES_STORE_ROOT + 'img_npy/' + identifier + 't_' + name.replace('/', '') + '_img',
tiles_img[i, j].astype(np.uint8))
asdf = np.load(TILES_STORE_ROOT + 'img_npy/' + identifier + 't_' + name.replace('/', '') + '_img.npy')
lab=labels_predicted,
alpha=0.5,
plot_img=True,
store_img=False,
plot_with='pyplot',
dir_name=RESULTS_ROOT,
file_name='test_image_x_' + area_name + RUN_NAME + str(number_of_training_steps),
cmap=cmap1)
img = lib.rgb_to_bgr(img)
blue_mean = channel_means['blue_mean']
green_mean = channel_means['green_mean']
red_mean = channel_means['red_mean']
img = lib.subtract_bgr_channel_means(img, blue_mean, green_mean, red_mean)
tiles_of_image = lib.image_to_tiles(img=img,
hgh=500,
wdt=500)
net.predict([tiles_of_image[0, 0, :, :, :]], oversample=False)
caffe_out = net.blobs['score'].data[0, :, :, :]
asdf = np.argmax(caffe_out, 0).astype(np.uint8)
lib.plot_sat_image_and_labels(img=img_cropped,
lab=asdf,
alpha=0.5,
plot_img=False,
store_img=True,
plot_with='Image',
dir_name=RESULTS_ROOT,
file_name='test_image_xx_' + area_name + RUN_NAME + str(number_of_training_steps),
cmap=cmap1)
img_ori = Image.open(image_file) # Cast image to numpy.ndarray. img = np.array(img_ori) # Bring image into required FCN input format. This is, swap RGB to BGR and subtract corresponding # per-channel means. This per-channel means can be used for the FCN models: # FCN-32s_PASCAL_Context, FCN-16s_PASCAL_Context & FCN-8s_PASCAL_Context img = lib.rgb_to_bgr(img) blue_mean = channel_means['blue_mean'] green_mean = channel_means['green_mean'] red_mean = channel_means['red_mean'] img = lib.subtract_bgr_channel_means(img, blue_mean, green_mean, red_mean) # Partition image into tiles. tiles_of_image = lib.image_to_tiles(img, tile_hgh, tile_wdt) tiles_of_image.shape # ---------------------------------------------------------------------------------------------------------------------- # Classify using pre-trained FCN. Note that it's unfortunately not possible to outsource the following 60 or so line # of codes into an own function. When this is done and the function is called on the remote python interpreter on Jan's # machine there is a problem. The problem is related to a pointer which is needed in the process of deconstructing the # caffe objects from the GPU, which is somehow not accessible and this causes the remote python interpreter to crash. # We don't want that, so the function code stands basically bellow. # ---------------------------------------------------------------------------------------------------------------------- # Classify using pre-trained FCN. data = tiles_of_image caffe_root = CAFFE_ROOT model_root = MODEL_ROOT caffe_mod = CAFFE_MODEL caffe_mod_par = CAFFE_MODEL_PAR