""" Author : amanbh - Set up some basic functions to load/manipulate image data - Visualize/Summarize cType counts, training data, and true classes - Plot Polygons with holes correctly by using descartes package Based on Kernel by Author : Oleg Medvedev Link : https://www.kaggle.com/torrinos/dstl-satellite-imagery-feature-detection/exploration-and-plotting/run/553107 """ import pandas as pd import numpy as np from shapely.wkt import loads as wkt_loads from matplotlib.patches import Polygon, Patch # decartes package makes plotting with holes much easier from descartes.patch import PolygonPatch import matplotlib.pyplot as plt import tifffile as tiff import pylab # turn interactive mode on so that plots immediately # See: http://stackoverflow.com/questions/2130913/no-plot-window-in-matplotlib # pylab.ion() inDir = '../input' # Give short names, sensible colors and zorders to object types CLASSES = { 1 : 'Bldg', 2 : 'Struct', 3 : 'Road', 4 : 'Track', 5 : 'Trees', 6 : 'Crops', 7 : 'Fast H20', 8 : 'Slow H20', 9 : 'Truck', 10 : 'Car', } COLORS = { 1 : '0.7', 2 : '0.4', 3 : '#b35806', 4 : '#dfc27d', 5 : '#1b7837', 6 : '#a6dba0', 7 : '#74add1', 8 : '#4575b4', 9 : '#f46d43', 10: '#d73027', } ZORDER = { 1 : 5, 2 : 5, 3 : 4, 4 : 1, 5 : 3, 6 : 2, 7 : 7, 8 : 8, 9 : 9, 10: 10, } # read the training data from train_wkt_v4.csv df = pd.read_csv(inDir + '/train_wkt_v4.csv') print(df.head()) # grid size will also be needed later.. gs = pd.read_csv(inDir + '/grid_sizes.csv', names=['ImageId', 'Xmax', 'Ymin'], skiprows=1) print(gs.head()) # imageIds in a DataFrame allImageIds = gs.ImageId.unique() trainImageIds = df.ImageId.unique() def get_image_names(imageId): ''' Get the names of the tiff files ''' d = {'3': '{}/three_band/{}.tif'.format(inDir, imageId), 'A': '{}/sixteen_band/{}_A.tif'.format(inDir, imageId), 'M': '{}/sixteen_band/{}_M.tif'.format(inDir, imageId), 'P': '{}/sixteen_band/{}_P.tif'.format(inDir, imageId), } return d def get_images(imageId, img_key = None): ''' Load images correspoding to imageId Parameters ---------- imageId : str imageId as used in grid_size.csv img_key : {None, '3', 'A', 'M', 'P'}, optional Specify this to load single image None loads all images and returns in a dict '3' loads image from three_band/ 'A' loads '_A' image from sixteen_band/ 'M' loads '_M' image from sixteen_band/ 'P' loads '_P' image from sixteen_band/ Returns ------- images : dict A dict of image data from TIFF files as numpy array ''' img_names = get_image_names(imageId) images = dict() if img_key is None: for k in img_names.keys(): images[k] = tiff.imread(img_names[k]) else: images[img_key] = tiff.imread(img_names[img_key]) return images def get_size(imageId): """ Get the grid size of the image Parameters ---------- imageId : str imageId as used in grid_size.csv """ xmax, ymin = gs[gs.ImageId == imageId].iloc[0,1:].astype(float) W, H = get_images(imageId, '3')['3'].shape[1:] return (xmax, ymin, W, H) def is_training_image(imageId): ''' Returns ------- is_training_image : bool True if imageId belongs to training data ''' return any(trainImageIds == imageId) def plot_polygons(fig, ax, polygonsList): ''' Plot descrates.PolygonPatch from list of polygons objs for each CLASS ''' legend_patches = [] for cType in polygonsList: print('{} : {} \tcount = {}'.format(cType, CLASSES[cType], len(polygonsList[cType]))) legend_patches.append(Patch(color=COLORS[cType], label='{} ({})'.format(CLASSES[cType], len(polygonsList[cType])))) for polygon in polygonsList[cType]: mpl_poly = PolygonPatch(polygon, color=COLORS[cType], lw=0, alpha=0.7, zorder=ZORDER[cType]) ax.add_patch(mpl_poly) # ax.relim() ax.autoscale_view() ax.set_title('Objects') ax.set_xticks([]) ax.set_yticks([]) return legend_patches def plot_image(fig, ax, imageId, img_key, selected_channels=None): ''' Plot get_images(imageId)[image_key] on axis/fig Optional: select which channels of the image are used (used for sixteen_band/ images) Parameters ---------- img_key : str, {'3', 'P', 'N', 'A'} See get_images for description. ''' images = get_images(imageId, img_key) img = images[img_key] title_suffix = '' if selected_channels is not None: img = img[selected_channels] title_suffix = ' (' + ','.join([ repr(i) for i in selected_channels ]) + ')' if len(img.shape) == 2: new_img = np.zeros((3, img.shape[0], img.shape[1])) new_img[0] = img new_img[1] = img new_img[2] = img img = new_img tiff.imshow(img, figure=fig, subplot=ax) ax.set_title(imageId + ' - ' + img_key + title_suffix) ax.set_xlabel(img.shape[-2]) ax.set_ylabel(img.shape[-1]) ax.set_xticks([]) ax.set_yticks([]) def visualize_image(imageId, plot_all=True): ''' Plot all images and object-polygons Parameters ---------- imageId : str imageId as used in grid_size.csv plot_all : bool, True by default If True, plots all images (from three_band/ and sixteen_band/) as subplots. Otherwise, only plots Polygons. ''' df_image = df[df.ImageId == imageId] xmax, ymin, W, H = get_size(imageId) if plot_all: fig, axArr = plt.subplots(figsize=(10, 10), nrows=3, ncols=3) ax = axArr[0][0] else: fig, axArr = plt.subplots(figsize=(10, 10)) ax = axArr if is_training_image(imageId): print('ImageId : {}'.format(imageId)) polygonsList = {} for cType in CLASSES.keys(): polygonsList[cType] = wkt_loads(df_image[df_image.ClassType == cType].MultipolygonWKT.values[0]) legend_patches = plot_polygons(fig, ax, polygonsList) ax.set_xlim(0, xmax) ax.set_ylim(ymin, 0) ax.set_xlabel(xmax) ax.set_ylabel(ymin) if plot_all: plot_image(fig, axArr[0][1], imageId, '3') plot_image(fig, axArr[0][2], imageId, 'P') plot_image(fig, axArr[1][0], imageId, 'A', [0, 3, 6]) plot_image(fig, axArr[1][1], imageId, 'A', [1, 4, 7]) plot_image(fig, axArr[1][2], imageId, 'A', [2, 5, 0]) plot_image(fig, axArr[2][0], imageId, 'M', [0, 3, 6]) plot_image(fig, axArr[2][1], imageId, 'M', [1, 4, 7]) plot_image(fig, axArr[2][2], imageId, 'M', [2, 5, 0]) if is_training_image(imageId): ax.legend(handles=legend_patches, # loc='upper center', bbox_to_anchor=(0.9, 1), bbox_transform=plt.gcf().transFigure, ncol=5, fontsize='x-small', title='Objects-' + imageId, # mode="expand", framealpha=0.3) return (fig, axArr, ax) # Loop over few training images and save to files for imageId in trainImageIds: fig, axArr, ax = visualize_image(imageId, plot_all=False) plt.savefig('Objects--' + imageId + '.png') plt.clf() # Optionally, view images immediately: # pylab.show() # Uncomment to show plot when interactive mode is off # (this function blocks till fig is closed)