def train_features():
datasets = load_pascal_set('train', False, 'aeroplane', 'bicycle', 'car', 'bird', 'boat', 'person')
train_set_x, train_set_y, train_images = datasets[0]
val_set_x, val_set_y, val_images = datasets[1]
all_descriptors = None
print('extract descriptors with SIFT')
desc_by_img = []
for img in train_set_x:
gray= cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
sift = cv2.SIFT()
kp, des = sift.detectAndCompute(gray,None)
desc_by_img.append(des)
if all_descriptors is None:
all_descriptors = des
else:
all_descriptors = np.concatenate([all_descriptors, des])
print('Find vocabulary with K-means clustering')
#pickle.dump(all_descriptors, open('all_descriptors.npy', 'wb'))
#pickle.dump(desc_by_img, open('desc_by_img.npy', 'wb'))
for nb_cluster in [8]:
model = KMeans(n_clusters=nb_cluster)
model.fit(all_descriptors)
print('Compute histos')
features = []
for descriptors, img_name, label in zip(desc_by_img, train_images, train_set_y):
feats = model.predict(descriptors)
histo = np.bincount(feats, minlength=nb_cluster)
feat_name_and_label = {}
feat_name_and_label['features'] = histo
feat_name_and_label['name'] = img_name
feat_name_and_label['label'] = label
features.append(feat_name_and_label)
f_features = open('output/pascal/'+str(nb_cluster)+'-cluster/train_features.npy', 'wb')
pickle.dump(features, f_features)
f_features.close()
f_model = open('output/pascal/'+str(nb_cluster)+'-cluster/train_kmeans_model.npy', 'wb')
pickle.dump(model, f_model)
f_model.close()
def bag_of_words():
datasets = load_pascal_set(False, 'aeroplane', 'bicycle', 'car', 'bird', 'boat', 'person')
train_set_x, train_set_y, train_images = datasets[0]
val_set_x, val_set_y, val_images = datasets[1]
all_descriptors = None
print('extract descriptors with SIFT')
desc_by_img = []
for img in train_set_x:
gray= cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
sift = cv2.SIFT()
kp, des = sift.detectAndCompute(gray,None)
desc_by_img.append(des)
if all_descriptors is None:
all_descriptors = des
else:
all_descriptors = np.concatenate([all_descriptors, des])
print('Find vocabulary with K-means clustering')
model = KMeans(n_clusters=settings.NB_CLUSTER)
model.fit(all_descriptors)
print('Compute histos')
features = []
for descriptors, img_name in zip(desc_by_img, train_images):
feats = model.predict(descriptors)
print(feats)
histo = np.bincount(feats, minlength=settings.NB_CLUSTER)
print(histo)
sys.exit('')
feat_and_name = {}
feat_and_name['features'] = histo
feat_and_name['name'] = img_name
features.append(feat_and_name)
pickle.dump(features, open('pascal_train_features.npy', 'wb'))
pickle.dump(model, open('pascal_train_model.npy', 'wb'))
def val_bag_of_words_prediction():
datasets = load_pascal_set('train', False, 'aeroplane', 'bicycle', 'car', 'bird', 'boat', 'person')
_, val_set_y, _ = datasets[1]
accuracies = []
nb_clusters = []
for nb_cluster in [32]:
val_features = pickle.load(open('output/pascal/'+str(nb_cluster)+'-cluster/val_features.npy', 'rb'))
clf = pickle.load(open('output/pascal/'+str(nb_cluster)+'-cluster/train_svm_classifier.npy', 'rb'))
y_pred = clf.predict(val_features)
accuracy = (y_pred == val_set_y).sum()/len(y_pred)
print("Accuracy of the model for nb clusters of %f is : %f" % (nb_cluster, accuracy))
accuracies.append(accuracy)
nb_clusters.append(nb_cluster)
pickle.dump(accuracies, open('output/pascal/accuracies.npy', 'wb'))
pickle.dump(nb_clusters, open('output/pascal/nb_clusters.npy', 'wb'))
def pascal_val_features(nb_cluster):
datasets = load_pascal_set('train', False, 'aeroplane', 'bicycle', 'car', 'bird', 'boat', 'person')
val_set_x, val_set_y, val_images = datasets[1]
model = pickle.load(open('output/pascal/'+str(nb_cluster)+'-cluster/train_kmeans_model.npy', 'rb'))
val_features = []
for img in val_set_x:
gray= cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
sift = cv2.SIFT()
kp, des = sift.detectAndCompute(gray,None)
features = model.predict(des)
histo = np.bincount(features, minlength=nb_cluster)
val_features.append(histo)
pickle.dump(val_features, open('output/pascal/'+str(nb_cluster)+'-cluster/val_features.npy', 'wb'))
