def test4():
dr = DataReader(*datareader_params)
image_segmenter = GenSeg(input_shape=[None, 176, 608, 3],
num_classes=num_classes,
load_model='saved/Long7-Lab-Fixed.ckpt')
velo_segmenter = GenSeg(input_shape=[None, 96, 96, 96, 1],
num_classes=num_classes,
load_model='saved/lidar.ckpt')
x_image = dr.get_image_data()
y_image_true = dr.get_image_labels()
n, _, _, = y_image_true.shape
image_hits = np.empty((3, 1))
image_totals = np.empty((3, 1))
for i in range(n):
print(i, 'image')
image_pred = np.argmax(image_segmenter.apply(x_image[i:i +
1, :, :, :]),
axis=-1)
image_true = y_image_true[i:i + 1, :, :]
for j in range(3, 6):
hit_pred = image_pred == j
hit_true = image_true == j
image_hits += np.sum(hit_true)
image_totals += np.sum(np.logical_and(hit_pred, hit_true))
x_velo = dr.get_velodyne_data()
y_velo_true = dr.get_velodyne_labels()
n, _, _, _ = y_velo_true.shape
velo_hits = np.empty((3, 1))
velo_totals = np.empty((3, 1))
for i in range(n):
print(i, 'velo')
velo_pred = np.argmax(velo_segmenter.apply(x_velo[i:i +
1, :, :, :, :]),
axis=-1)
velo_true = y_velo_true[i:i + 1, :, :, :]
for j in range(3, 6):
hit_pred = velo_pred == j
hit_true = velo_true == j
velo_hits += np.sum(hit_true)
velo_totals += np.sum(np.logical_and(hit_pred, hit_true))
for total in image_totals:
print(total)
for hit in image_hits:
print(hit)
for total in velo_totals:
print(total)
for hit in velo_hits:
print(hit)
def test3(name):
"""Calculate average accuracy"""
input_shape = [None, 176, 608, 3]
dr = DataReader(*datareader_params)
x = dr.get_image_data()
y = dr.get_image_labels()
func = np.vectorize(original_to_label)
y = func(y)
n, _, _, _ = x.shape
model = GenSeg(input_shape=input_shape, num_classes=num_classes, load_model=name)
average = 0
count = 0
for i in range(0,n,batch_size):
batch_data = x[i:i+batch_size, :, :, :]
batch_labels = y[i:i+batch_size, :, :]
results = model.apply(batch_data)
results = np.argmax(results, axis=-1)
results = np.equal(results, batch_labels)
results = np.average(results.astype(dtype=np.float32))
average += results
count += 1
print(average/count)
def test4(name):
"""Render results"""
rcParams['figure.figsize'] = 40, 15
input_shape = [None, 176, 608, 3]
dr = DataReader(*datareader_params)
x = dr.get_image_data()
y = dr.get_image_labels()
func = np.vectorize(original_to_label)
y = func(y)
n, _, _, _ = x.shape
model = GenSeg(input_shape=input_shape, num_classes=num_classes, load_model=name)
#fourcc = cv2.VideoWriter_fourcc(*'MJPG')
#out = cv2.VideoWriter('out.avi', fourcc, 20.0, tuple(input_shape[1:-1]))
f = None
for i in range(0, n, batch_size):
batch_data = x[i:i + batch_size, :, :, :]
batch_labels = y[i:i + batch_size, :, :]
results = model.apply(batch_data)
results = np.argmax(results, axis=-1)
for img in results:
colored = np.empty(input_shape[1:], dtype=np.uint8)
for (idx, idy), value in np.ndenumerate(img):
colored[idx, idy, :] = get_color(label_to_original(value))
if f is None:
f = plt.imshow(colored)
else:
f.set_data(colored)
plt.pause(.05)
plt.draw()
def test6(name):
dr = DataReader(*datareader_params)
x = dr.get_velodyne_data()
y = dr.get_velodyne_labels()
func = np.vectorize(original_to_label)
y = func(y)
n, _, _, _, _ = x.shape
batch_size = 1
model = GenSeg(input_shape=input_shape,
num_classes=num_classes,
load_model=name)
average = 0
count = 0
for i in range(0, n, batch_size):
batch_data = x[i:i + batch_size, ...]
batch_labels = y[i:i + batch_size, ...]
results = model.apply(batch_data)
results = np.argmax(results, axis=-1)
results = np.equal(results, batch_labels)
results = np.average(results.astype(dtype=np.float32))
average += results
count += 1
print(i)
print(average / count)
def test3(name):
dr = DataReader(*datareader_params)
x = dr.get_velodyne_data()
model = GenSeg(input_shape=input_shape,
num_classes=num_classes,
load_model=name)
datapoint = model.apply(x[:1, :, :, :, :])
datapoint = np.argwhere(datapoint == 3)
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.scatter(datapoint[:, 0], datapoint[:, 1], datapoint[:, 2])
plt.show()
def test2(name):
"""Test on selected testing images"""
filenames = [
'data/image_data/testing/0000/000000.png',
'data/image_data/testing/0000/000040.png',
'data/image_data/testing/0004/000000.png',
'data/image_data/testing/0005/000020.png',
'data/image_data/testing/0005/000240.png'
]
shape = (len(filenames), 176, 608, 3)
n, h, w, c = shape
image_data = np.zeros((n, h, w, c))
i = 0
for f in filenames:
image = normalize_img(imread(f)) # Fix brightness and convert to lab colorspace
image_data[i, :, :, :] = image[:h*2:2, :w*2:2, :]
'''plt.figure()
plt.imshow(image_data[i])
plt.figure()
plt.imshow(lab2rgb(image_data[i]))
plt.show()'''
i += 1
model = GenSeg(input_shape=[None, h, w, c], num_classes=num_classes, load_model=name)
result = model.apply(image_data)
result = np.argmax(result, axis=-1)
'''for img in result:
plt.figure()
plt.imshow(img.astype(np.uint8))
plt.show()'''
colored = np.empty(shape)
for (i, x, y), value in np.ndenumerate(result):
colored[i, x, y] = get_color(label_to_original(value))
i = 0
for img in colored:
img = img.astype(np.uint8, copy=False)
imsave('%d.png' % i, img, 'png')
'''plt.figure()
plt.imshow(img)
plt.show()'''
i += 1
def test5(name):
image_segmenter = GenSeg(input_shape=[None, 176, 608, 3],
num_classes=num_classes,
load_model='saved/Long7-Lab-Fixed.ckpt')
velo_segmenter = GenSeg(input_shape=[None, 96, 96, 96, 1],
num_classes=num_classes,
load_model='saved/lidar.ckpt')
dr = DataReader(*datareader_params)
images = dr.get_image_data()[:1, :, :, :]
image_preds = image_segmenter.apply(images)
image_pred = image_preds[0, :, :, :]
image_candidates = []
for i in range(3, 6):
class_hits = np.argmax(image_pred, axis=-1) == i
class_hits = np.argwhere(class_hits)
if len(class_hits) > 0:
confidences = image_pred[class_hits[:, 0], class_hits[:, 1], :]
dbscan = DBSCAN(eps=50, min_samples=1)
dbscan.fit(class_hits)
labels = dbscan.labels_
for j in range(np.max(labels) + 1):
idxs = labels == j
idxs = np.argwhere(idxs)
avg_conf = confidences[idxs, :]
avg_conf = np.mean(avg_conf, 0)
com = np.mean(class_hits[idxs, :], 0)
image_candidates.append((i, avg_conf, com))
velos = dr.get_velodyne_data()[:1, :, :, :, :]
velo_preds = velo_segmenter.apply(velos)
velo_pred = velo_preds[0, :, :, :, :]
velo_candidates = []
for i in range(3, 6):
class_hits = np.argmax(velo_pred, axis=-1) == i
class_hits = np.argwhere(class_hits)
if len(class_hits) > 0:
confidences = velo_pred[class_hits[:, 0], class_hits[:, 1],
class_hits[:, 2], :]
dbscan = DBSCAN(eps=50, min_samples=1)
dbscan.fit(class_hits)
labels = dbscan.labels_
for j in range(np.max(labels) + 1):
idxs = labels == j
idxs = np.argwhere(idxs)
avg_conf = confidences[idxs, :]
avg_conf = np.mean(avg_conf, 0)
com = np.mean(class_hits[idxs, :], 0)
velo_candidates.append((i, avg_conf, com))
print('IMAGE CANDIDATES')
for (class_number, avg_conf, com) in image_candidates:
print('Class number:', class_number)
print('Average confidence:')
for conf in avg_conf:
print(conf)
print('Center of mass:')
for coord in com:
print(coord)
print('')
print('VELODYNE CANDIDATES')
for (class_number, avg_conf, com) in velo_candidates:
print('Class number:', class_number)
print('Average confidence:')
for conf in avg_conf:
print(conf)
print('Center of mass:')
for coord in com:
print(coord)
print('')