Eventually found the problem - I was using NormalBayesClassifier wrong. It isn't meant to be fed dozens of HD images directly: one should first munge them using OpenCV's other algorithms.
I have ended up doing the following:
+ Crop the image down to an area which may contain the object
+ Turn image to greyscale
+ Use cv2.goodFeaturesToTrack() to collect features from the cropped area to train the classifier.
A tiny number of features works for me, perhaps because I've cropped the image right down and its fortunate enough to contain high-contrast objects that get obscured for one class.
The following code gets as much as 95% of the population correct:
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import cv2
import sys, os.path, getopt
import numpy, random
def _usage():
print
print "cvbayes trainer"
print
print "Options:"
print
print "-m --ham= path to dir of ham images"
print "-s --spam= path to dir of spam images"
print "-h --help this help text"
print "-v --verbose lots more output"
print
def _parseOpts(argv):
"""
Turn options + args into a dict of config we'll follow. Merge in default conf.
"""
try:
opts, args = getopt.getopt(argv[1:], "hm:s:v", ["help", "ham=", 'spam=', 'verbose'])
except getopt.GetoptError as err:
print(err) # will print something like "option -a not recognized"
_usage()
sys.exit(2)
optsDict = {}
for o, a in opts:
if o == "-v":
optsDict['verbose'] = True
elif o in ("-h", "--help"):
_usage()
sys.exit()
elif o in ("-m", "--ham"):
optsDict['ham'] = a
elif o in ('-s', '--spam'):
optsDict['spam'] = a
else:
assert False, "unhandled option"
for mandatory_arg in ('ham', 'spam'):
if mandatory_arg not in optsDict:
print "Mandatory argument '%s' was missing; cannot continue" % mandatory_arg
sys.exit(0)
return optsDict
class ClassifierWrapper(object):
"""
Setup and encapsulate a naive bayes classifier based on OpenCV's
NormalBayesClassifier. Presently we do not use it intelligently,
instead feeding in flattened arrays of B&W pixels.
"""
def __init__(self):
super(ClassifierWrapper,self).__init__()
self.classifier = cv2.NormalBayesClassifier()
self.data = []
self.responses = []
def _load_image_features(self, f):
image_colour = cv2.imread(f)
image_crop = image_colour[327:390, 784:926] # Use the junction boxes, luke
image_grey = cv2.cvtColor(image_crop, cv2.COLOR_BGR2GRAY)
features = cv2.goodFeaturesToTrack(image_grey, 4, 0.02, 3)
return features.flatten()
def train_from_file(self, f, cl):
features = self._load_image_features(f)
self.data.append(features)
self.responses.append(cl)
def train(self, update=False):
matrix_data = numpy.matrix( self.data ).astype('float32')
matrix_resp = numpy.matrix( self.responses ).astype('float32')
self.classifier.train(matrix_data, matrix_resp, update=update)
self.data = []
self.responses = []
def predict_from_file(self, f):
features = self._load_image_features(f)
features_matrix = numpy.matrix( [ features ] ).astype('float32')
retval, results = self.classifier.predict( features_matrix )
return results
if __name__ == "__main__":
opts = _parseOpts(sys.argv)
cw = ClassifierWrapper()
ham = os.listdir(opts['ham'])
spam = os.listdir(opts['spam'])
n_training_samples = min( [len(ham),len(spam)])
print "Will train on %d samples for equal sets" % n_training_samples
for f in random.sample(ham, n_training_samples):
img_path = os.path.join(opts['ham'], f)
print "ham: %s" % img_path
cw.train_from_file(img_path, 2)
for f in random.sample(spam, n_training_samples):
img_path = os.path.join(opts['spam'], f)
print "spam: %s" % img_path
cw.train_from_file(img_path, 1)
cw.train()
print
print
# spam dir much bigger so mostly unused, let's try predict() on all of it
print "predicting on all spam..."
n_wrong = 0
n_files = len(os.listdir(opts['spam']))
for f in os.listdir(opts['spam']):
img_path = os.path.join(opts['spam'], f)
result = cw.predict_from_file(img_path)
print "%s\t%s" % (result, img_path)
if result[0][0] == 2:
n_wrong += 1
print
print "got %d of %d wrong = %.1f%%" % (n_wrong, n_files, float(n_wrong)/n_files * 100, )
Right now I'm training it with a random subset of the spam, simply because there's much more of it and you should have a roughly equal amount of training data for each class. With better curated data (e.g. always include samples from dawn and dusk when lighting is different) it would probably be higher.
Perhaps even a NormalBayesClassifier is the wrong tool for the job and I should experiment with motion detection across consecutive frames - but at least the Internet has an example to pick apart now.
