almost full flow

05_build_samplesfile.py

@@ -0,0 +1,44 @@
+from utils.geo import get_rectangle
+from utils.geo import WAYS_DATA_FILENAME
+import json
+import os
+
+# We need the elements
+print 'Loading %s...' % WAYS_DATA_FILENAME
+with open(WAYS_DATA_FILENAME, 'r') as f:
+    ways_data = json.load(f)
+
+samples_data = {}
+
+image_files = [f for f in os.listdir('satellite/gray') if f.endswith('.png')]
+for image_file in image_files:
+    print 'Processing %s...' % image_file
+
+    # Find the category
+    sport = image_file[image_file.find('_')+1:image_file.rfind('_')]
+    if sport not in samples_data:
+        samples_data[sport] = []
+
+    # The ID is between the last underscore and the extension dot
+    # For example: pitch_volleyball_268478401.png -> 268478401
+    way_id = image_file[image_file.rfind('_') + 1:image_file.find('.')]
+    bounds = ways_data[way_id]['bounds']
+
+    # Add a rectangle with the feature
+    x, y, w, h = get_rectangle(bounds=bounds)
+    if w <= 25 or h <= 25:
+        print 'Pic not added'
+        continue
+    if x <= 0 or y <= 0 or w <= 0 or h <= 0:
+        print 'Pic not added'
+        continue
+    entry = 'satellite/gray/%s\t1\t%d\t%d\t%d\t%d\n' % (image_file, x, y, w, h)
+    samples_data[sport].append(entry)
+
+for sport in samples_data.keys():
+    datafile = 'info_%s.dat' % sport
+    print 'Saving data file: %s' % datafile
+    with open(datafile, 'w') as f:
+        print len(samples_data[sport])
+        for entry in samples_data[sport]:
+            f.write(entry)

05_draw_bbox.py

@@ -3,74 +3,38 @@
 will use this box to better train our ML algo.
 '''
 
+from utils.geo import get_rectangle
 from utils.geo import WAYS_DATA_FILENAME
 import cv2
 import json
-import math
 import os
 
-image_width = 1280
-image_height = 1280
-
-def get_rectangle(bounds):
-    # This converts a latitude delta into an image delta. For USA, at zoom
-    # level 19, we know that we have 0.21 meters/pixel. So, an image is showing
-    # about 1280 pixels * 0.21 meters/pixel = 268.8 meters.
-    # On the other hand we know that at the same angle, a degress in latlon is:
-    # https://en.wikipedia.org/wiki/Latitude
-    # latitude = 111,132 m
-    # longitude = 78,847 m
-    # Finally, we add 5% to make sure the feature is contained within the
-    # rectangle (from what we see, the bbox are tight around the feature)
-    latitude_factor  = 1.05 * 111132.0 / 0.21
-    longitude_factor = 1.05 * 78847.0 / 0.21
-
-    # Feature size
-    feature_width = longitude_factor * math.fabs(bounds[1] - bounds[3])
-    feature_height = latitude_factor * math.fabs(bounds[0] - bounds[2])
-    if feature_width > image_width or feature_height > image_height:
-        print '** Warning ** The feature is bigger than the image.'
-
-    # CV params (int required)
-    x = int((image_width / 2) - (feature_width / 2))
-    y = int((image_height / 2) - (feature_height / 2))
-    w = int(feature_width)
-    h = int(feature_height)
-    if w <= 25 or h <= 25:
-        print '** Warning ** This image has very narrow bounds.'
-        print bounds
-        print x, y, w, h
-    if x == 0 or y == 0 or w == 0 or h == 0:
-        print '** Warning ** There is something wrong with this image bounds.'
-        print bounds
-        print x, y, w, h
-    return x, y, w, h
-
 # We need the elements
-print 'Loading %s...' % WAYS_DATA_FILENAME
-with open(WAYS_DATA_FILENAME, 'r') as f:
-    ways_data = json.load(f)
+# print 'Loading %s...' % WAYS_DATA_FILENAME
+# with open(WAYS_DATA_FILENAME, 'r') as f:
+#     ways_data = json.load(f)
 
 image_files = [f for f in os.listdir('satellite') if f.endswith('.png')]
+print len(image_files)
 
 for image_file in image_files:
     print 'Processing %s...' % image_file
     # The ID is between the last underscore and the extension dot
     # For example: pitch_volleyball_268478401.png -> 268478401
-    way_id = image_file[image_file.rfind('_') + 1:image_file.find('.')]
-    bounds = ways_data[way_id]['bounds']
+    # way_id = image_file[image_file.rfind('_') + 1:image_file.find('.')]
+    # bounds = ways_data[way_id]['bounds']
 
     # Add a rectangle with the feature
-    x, y, w, h = get_rectangle(bounds=bounds)
-    img = cv2.imread(os.path.join('satellite', image_file))
-    cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
-    cv2.imwrite(os.path.join('satellite/rectangle', image_file), img)
+    # x, y, w, h = get_rectangle(bounds=bounds)
+    # img = cv2.imread(os.path.join('satellite', image_file))
+    # cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
+    # cv2.imwrite(os.path.join('satellite/rectangle', image_file), img)
+
+    # To show the image
+    # cv2.imshow('img',img)
+    # cv2.waitKey(0)
+    # cv2.destroyAllWindows()
 
     # Generate a grayscale version
     img_gray = cv2.imread(os.path.join('satellite', image_file), 0)
     cv2.imwrite(os.path.join('satellite/gray', image_file), img_gray)
-
-# To show the image
-# cv2.imshow('img',img)
-# cv2.waitKey(0)
-# cv2.destroyAllWindows()

07_fit_min_neighbors.py

@@ -0,0 +1,62 @@
+from __future__ import division
+import csv
+import cv2
+import numpy as np
+import os
+
+tennis_cascade_files = [
+    'output/cascade-default.xml',
+    'output/cascade-4000-2000.xml',
+    'output/cascade-6000-3000.xml',
+    'output/cascade-8000-4000.xml']
+
+positive_files = [os.path.join('satellite/fit', f) \
+    for f in os.listdir('satellite/fit') if f.endswith('.png')]
+
+negative_files = [os.path.join('satellite/fit/negative', f) \
+    for f in os.listdir('satellite/fit/negative') if f.endswith('.png')]
+
+def get_total_pitches(tennis_cascade, filename, min_neighbors):
+    img = cv2.imread(filename, 0)
+    pitches = tennis_cascade.detectMultiScale(
+        img, minNeighbors=min_neighbors)
+    return len(pitches)
+
+for tennis_cascade_file in tennis_cascade_files:
+    print tennis_cascade_file
+    tennis_cascade = cv2.CascadeClassifier(tennis_cascade_file)
+
+    # Open
+    positive_f = open(tennis_cascade[:-4] + '_positive.csv', 'w')
+    negative_f = open(tennis_cascade[:-4] + '_negative.csv', 'w')
+    positive_writer = csv.writer(positive_f)
+    negative_writer = csv.writer(negative_f)
+
+    for min_neighbors in range(0, 501, 10):
+        print min_neighbors
+
+        # Pos
+        total_set = 0
+        for positive_file in positive_files:
+            total_pitches = get_total_pitches(
+                tennis_cascade=tennis_cascade,
+                filename=positive_file,
+                min_neighbors=min_neighbors)
+            total_set += total_pitches
+        total_average = total_set / len(positive_files)
+        positive_writer.writerow([total_average, min_neighbors])
+
+        # Neg
+        total_set = 0
+        for negative_file in negative_files:
+            total_pitches = get_total_pitches(
+                tennis_cascade=tennis_cascade,
+                filename=negative_file,
+                min_neighbors=min_neighbors)
+            total_set += total_pitches
+        total_average = total_set / len(negative_files)
+        negative_writer.writerow([total_average, min_neighbors])
+
+    # Close
+    positive_f.close()
+    negative_f.close()

Makefile

@@ -25,6 +25,8 @@ step4:
 	python 04_download_satellite.py --sport volleyball --count 10
 	python 04_download_satellite.py --sport multi --count 10
 	python 04_download_satellite.py --sport softball --count 10
+	# mv satellite/*png satellite/training/
+	# mv satellite/*png satellite/testing/
 
 step5:
 	# Draw the bbox on the test files
@@ -33,3 +35,13 @@ step5:
 step6:
 	# Get some random images
 	python 06_get_negatives.py --count 25
+	find satellite/negative -type f > negative.txt
+
+createsamples:
+	opencv_createsamples -info info_baseball.dat -num 99 -vec info_baseball.vec
+	opencv_createsamples -info info_basketball.dat -num 100 -vec info_basketball.vec
+	opencv_createsamples -info info_tennis.dat -num 9998 -vec info_tennis.vec
+	opencv_traincascade -data output -vec info_baseball.vec -bg negative.txt -numPos 99 -numNeg 100
+	opencv_traincascade -data output -vec info_basketball.vec -bg negative.txt -numPos 100 -numNeg 100
+	opencv_traincascade -data output -vec info_tennis.vec -bg negative.txt -numPos 9998 -numNeg 5000
+

detector.py

@@ -0,0 +1,58 @@
+import numpy as np
+import cv2
+
+baseball_cascade = cv2.CascadeClassifier('baseball.xml')
+basketball_cascade = cv2.CascadeClassifier('basketball.xml')
+tennis_cascade = cv2.CascadeClassifier('tennis.xml')
+
+# testfile = 'satellite/training/pitch_baseball_220727025.png'
+# testfile = 'satellite/training/pitch_baseball_222703638.png'
+# testfile = 'satellite/training/pitch_baseball_223914194.png'
+# testfile = 'satellite/training/pitch_baseball_226905824.png'
+# testfile = 'satellite/training/pitch_baseball_227372226.png'
+# testfile = 'satellite/training/pitch_baseball_227683244.png'
+
+# testfile = 'satellite/detection/pitch_baseball_133230978.png'
+# testfile = 'satellite/detection/pitch_baseball_133593974.png'
+# testfile = 'satellite/detection/pitch_baseball_134874855.png'
+# testfile = 'satellite/detection/pitch_baseball_199130202.png'
+# testfile = 'satellite/detection/pitch_baseball_284527697.png'
+# testfile = 'satellite/detection/pitch_baseball_317137177.png'
+# testfile = 'satellite/detection/pitch_baseball_48331085.png'
+# testfile = 'satellite/detection/pitch_baseball_68467399.png'
+# testfile = 'satellite/detection/pitch_baseball_81189377.png'
+# testfile = 'satellite/detection/pitch_baseball_97575184.png'
+
+# testfile = 'satellite/detection/pitch_tennis_105660674.png'
+# testfile = 'satellite/detection/pitch_tennis_120231577.png'
+# testfile = 'satellite/detection/pitch_tennis_172547292.png'
+# testfile = 'satellite/detection/pitch_tennis_177425633.png'
+# testfile = 'satellite/detection/pitch_tennis_224740547.png'
+# testfile = 'satellite/detection/pitch_tennis_250911604.png'
+# testfile = 'satellite/detection/pitch_tennis_285058169.png'
+# testfile = 'satellite/detection/pitch_tennis_290182837.png'
+# testfile = 'satellite/detection/pitch_tennis_302813940.png'
+# testfile = 'satellite/detection/pitch_tennis_343232913.png'
+
+# testfile = 'satellite/detection/pitch_basketball_139165791.png'
+# testfile = 'satellite/detection/pitch_basketball_156416713.png'
+testfile = 'satellite/detection/pitch_basketball_242894925.png'
+# testfile = 'satellite/detection/pitch_basketball_256427642.png'
+# testfile = 'satellite/detection/pitch_basketball_271825251.png'
+# testfile = 'satellite/detection/pitch_basketball_276916820.png'
+# testfile = 'satellite/detection/pitch_basketball_302422677.png'
+# testfile = 'satellite/detection/pitch_basketball_331162643.png'
+# testfile = 'satellite/detection/pitch_basketball_332627356.png'
+# testfile = 'satellite/detection/pitch_basketball_48665083.png'
+
+img = cv2.imread(testfile)
+gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+
+pitches = basketball_cascade.detectMultiScale(gray, minNeighbors=200)
+print 'Pitches found: %d' % len(pitches)
+for (x,y,w,h) in pitches:
+    cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
+
+cv2.imshow('img', img)
+cv2.waitKey(0)
+cv2.destroyAllWindows()

satellite/.gitignore

@@ -1,2 +1,3 @@
 gray
 rectangle
+fit

utils/geo.py

@@ -1,4 +1,8 @@
 from shapely.geometry import Polygon
+import math
+
+image_width = 1280
+image_height = 1280
 
 ELEMENTS_FILENAME = 'data/elements.json'
 WAYS_DATA_FILENAME = 'data/ways.json'
@@ -27,3 +31,35 @@ def get_ways_data(elements, coords):
 
     # Done
     return ways_data
+
+def get_rectangle(bounds):
+    # This converts a latitude delta into an image delta. For USA, at zoom
+    # level 19, we know that we have 0.21 meters/pixel. So, an image is showing
+    # about 1280 pixels * 0.21 meters/pixel = 268.8 meters.
+    # On the other hand we know that at the same angle, a degress in latlon is:
+    # https://en.wikipedia.org/wiki/Latitude
+    # latitude = 111,132 m
+    # longitude = 78,847 m
+    latitude_factor  = 111132.0 / 0.21
+    longitude_factor = 78847.0 / 0.21
+
+    # Feature size
+    feature_width = longitude_factor * math.fabs(bounds[1] - bounds[3])
+    feature_height = latitude_factor * math.fabs(bounds[0] - bounds[2])
+    if feature_width > image_width or feature_height > image_height:
+        print '** Warning ** The feature is bigger than the image.'
+
+    # CV params (int required)
+    x = int((image_width / 2) - (feature_width / 2))
+    y = int((image_height / 2) - (feature_height / 2))
+    w = int(feature_width)
+    h = int(feature_height)
+    if w <= 25 or h <= 25:
+        print '** Warning ** This image has very narrow bounds.'
+        print bounds
+        print x, y, w, h
+    if x <= 0 or y <= 0 or w <= 0 or h <= 0:
+        print '** Warning ** There is something wrong with this image bounds.'
+        print bounds
+        print x, y, w, h
+    return x, y, w, h