def get_actual_odom(image1_path,image2_path):
'''Get actual relative odometry.'''
image1_seq_id, image1_frame=parse_path(image1_path)
image2_seq_id, image2_frame=parse_path(image2_path)
image1_odom=read_odom(sequence_id=image1_seq_id,desired_frame=image1_frame)
image2_odom=read_odom(sequence_id=image2_seq_id,desired_frame=image2_frame)
odom_dt_actual=image1_odom-image2_odom
return odom_dt_actual
def _valBatchGenerator(X_val_filelist, y_val_filelist, batchSize):
"""
Yield X and Y data when the batch is filled.
"""
#Sort filelists to confirm they are same order
X_val_filelist.sort(key=lambda f: int(''.join(filter(str.isdigit, f))))
y_val_filelist.sort(key=lambda f: int(''.join(filter(str.isdigit, f))))
#Do not shuffle!
while True:
idx = 1 #Skip first image since we need t-1
while idx + (batchSize - 1) < len(X_val_filelist):
X_val_1 = np.zeros((batchSize, 192, 640, 3),
dtype=np.uint8) #time=t
X_val_2 = np.zeros((batchSize, 192, 640, 3),
dtype=np.uint8) #time=t-1
# y_val_depth=np.zeros((batchSize,192,640),dtype=np.uint8) #time=t depth
y_val_odom = np.zeros((batchSize, 6), dtype=np.float64) #dt odom
y_val_rpy = np.zeros((batchSize, 3), dtype=np.float64)
y_val_xyz = np.zeros((batchSize, 3), dtype=np.float64)
y_val_rpyxyz = np.zeros((batchSize, 6), dtype=np.float64)
for i in range(batchSize):
#Load images
X_val_1[i] = rgb_read(X_val_filelist[idx + i]) #time=t
X_val_2[i] = rgb_read(X_val_filelist[idx - 1 + i]) #time=t-1
# y_val_depth[i]=depth_read(y_val_filelist[idx+i]) #time=t
#Calculate change in odometry
current_filename = X_val_filelist[idx + i] #time=t
prev_filename = X_val_filelist[idx - 1 + i] #time=t-1
sequence_id, frame_id = basename(current_filename).split(
'_sync_')
prev_sequence_id, prev_frame_id = basename(
prev_filename).split('_sync_')
#print('original:' + frame_id)
frame_id = int(frame_id.split('.')[0])
prev_frame_id = int(prev_frame_id.split('.')[0])
#print('converted:' + str(frame_id))
#prev_frame_id=frame_id-1
current_odom = read_odom(sequence_id, frame_id)
prev_odom = read_odom(prev_sequence_id, prev_frame_id)
#print('Val: '+f'{frame_id}, {prev_frame_id}')
y_val_odom[i] = normalize(current_odom - prev_odom)
# y_val_odom[i]=current_odom-prev_odom
y_val_rpy[i] = y_val_odom[i][0:3]
y_val_xyz[i] = y_val_odom[i][3:6]
y_val_rpyxyz[i] = y_val_odom[i]
#Reshape [samples][width][height][pixels]
X_val_1 = X_val_1.reshape(X_val_1.shape[0], X_val_1.shape[1],
X_val_1.shape[2],
X_val_1.shape[3]).astype(np.uint8)
X_val_2 = X_val_2.reshape(X_val_2.shape[0], X_val_2.shape[1],
X_val_2.shape[2],
X_val_2.shape[3]).astype(np.uint8)
# y_val_depth = y_val_depth.reshape((y_val_depth.shape[0],1,-1)).astype(np.uint8)
# y_val_depth = y_val_depth.squeeze()
# normalize inputs and outputs from 0-255 to 0-1
X_val_1 = np.divide(X_val_1, 255).astype(np.float16)
X_val_2 = np.divide(X_val_2, 255).astype(np.float16)
# y_val_depth=np.divide(y_val_depth,255).astype(np.float16)
# if (idx % 68)==0: #1024
# print(str(idx)+'/'+str(len(X_val_filelist)))
idx += batchSize
#y_val_depth=y_val_depth.reshape((batchSize,len(y_val_depth)))
# y_val_depth=y_val_depth.reshape((batchSize,y_val_depth[0].size))
#Provide both images [time=t, time(t-1)]
X_val = [X_val_1, X_val_2]
#Provide depth and odom
y_val = [y_val_rpy, y_val_xyz]
# y_val=[y_val_rpyxyz]
# print('Val:', y_val)
#print('Val: '+str(y_val_depth.shape)+','+str(y_val_odom.shape))
yield X_val, y_val
if __name__ == '__main__':
model = models.mock_undeepvo_withflow
model_name = 'mock_undeepvo_withflow'
model = main(model_name=model_name,
model=model,
num_epochs=80,
batch_size=32)
show_test_image = True
if show_test_image:
#"data\train\X\2011_09_30_drive_0016_sync_0000000006.png"
#"data\val\X\2011_09_30_drive_0018_sync_0000000135.png"
image1_path = r"data\val\X\2011_09_30_drive_0018_sync_0000000135.png"
image2_path = r"data\val\X\2011_09_30_drive_0018_sync_0000000134.png"
flow_path = r"data\val\flow\2011_09_30_drive_0018_sync_0000000135_flow.png"
image1_odom = read_odom(sequence_id="2011_09_30_drive_0018",
desired_frame=135)
image2_odom = read_odom(sequence_id="2011_09_30_drive_0018",
desired_frame=134)
#Read test image
image1 = rgb_read(image1_path) #640x480, 1242x375
image1 = image1.reshape(1, 192, 640, 3)
image1 = np.divide(image1, 255).astype(np.float16)
#Read test image
image2 = rgb_read(image2_path) #640x480
image2 = image2.reshape(1, 192, 640, 3)
image2 = np.divide(image2, 255).astype(np.float16)
#Read test image
flow = rgb_read(flow_path) #640x480
