def canonization_training(opts):
print "# Canonization training"
params = caching.nul_repr_dict()
# Generate average intensity image from a subset of the dataset.
img_avg = np.zeros(opts["img_shape"], dtype=int)
files = dataset.training_files(opts["num_train_images"])
for img_file, depth_file in print_progress(files):
img = preprocess_image(img_file)
segmask = segmentation(depth_file, opts["segmentation"])
img, segmask = _canonize(img, segmask, opts)
# Orient correctly if image is upside down
if dataset.is_upside_down(img_file):
img = np.fliplr(np.flipud(img))
img_avg += img
img_avg /= len(files)
params["img_avg"] = img_avg
params["img_avg_upsidedown"] = np.fliplr(np.flipud(img_avg))
return params
def bow_train(featfunc, opts, canon_opts, params):
descs = []
files = dataset.training_files(opts['num_train_images'])
for img_file, depth_file in print_progress(files):
img, segmask = canonize(img_file, depth_file, canon_opts, params)
if featfunc == 'daisy':
h = daisy(img_as_float(img), **opts['feature_opts'])
if featfunc == 'jet':
h = jet(img_as_float(img), **opts['feature_opts'])
segmask = sp.misc.imresize(segmask, h.shape[:2])
for i in range(h.shape[0]):
for j in range(h.shape[1]):
if segmask[i,j] != 0:
descs.append(h[i,j,:].flatten())
descs = [descs[i] for i in range(0, len(descs), opts['feature_step'])]
descs = np.vstack(tuple(descs))
print '# K-means clustering of %i features of dimensionality %i'%(descs.shape[0], descs.shape[1])
kmeans = KMeans(opts['num_clusters'], n_jobs=options.num_threads)
kmeans.fit(descs)
return kmeans
