def get_experiment_info(checkpoint_name):
exp_output_base_dir = core.data_dir() + '/outputs/' + checkpoint_name
# Parse experiment config
config_file = exp_output_base_dir + '/{}.yaml'.format(checkpoint_name)
config = config_utils.parse_yaml_config(config_file)
predictions_base_dir = exp_output_base_dir + '/predictions'
return config, predictions_base_dir
def parse_yaml_config(yaml_path):
"""Parses a yaml config
Args:
yaml_path: path to yaml config
Returns:
config_obj: config converted to object
"""
# Add check for duplicate keys in yaml
yaml.add_constructor(yaml.resolver.BaseResolver.DEFAULT_MAPPING_TAG,
no_duplicates_constructor)
with open(yaml_path, 'r') as yaml_file:
config_dict = yaml.load(yaml_file)
config_obj = config_dict_to_object(config_dict)
config_obj.config_name = os.path.splitext(os.path.basename(yaml_path))[0]
config_obj.exp_output_dir = core.data_dir(
) + '/outputs/' + config_obj.config_name
# Prepend data folder to paths
paths_config = config_obj.train_config.paths_config
if paths_config.checkpoint_dir is None:
checkpoint_dir = config_obj.exp_output_dir + '/checkpoints'
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
paths_config.checkpoint_dir = checkpoint_dir
else:
paths_config.checkpoint_dir = os.path.expanduser(
paths_config.checkpoint_dir)
paths_config.logdir = config_obj.exp_output_dir + '/logs'
paths_config.pred_dir = config_obj.exp_output_dir + '/predictions'
return config_obj
def main():
#########################################################
# Specify Source Folders and Parameters For Frame Reader
#########################################################
data_split_dir = 'training'
# Specify whether the validation or inference results need to be
# visualized.
#results_dir = 'validation'
results_dir = 'testing'
# sample_free, anchor_redundancy, black_box, bayes_od_none,
# bayes_od_ci_fast. bayes_od_ci,or bayes_od_ici
uncertainty_method = 'bayes_od'
dataset_dir = os.path.expanduser('~/Datasets/Kitti/object/')
image_dir = os.path.join(dataset_dir, data_split_dir) + '/image_2'
label_dir = os.path.join(dataset_dir, data_split_dir) + '/label_2'
checkpoint_name = 'retinanet_bdd'
checkpoint_number = '101'
if results_dir == 'testing':
prediction_dir = os.path.join(core.data_dir(), 'outputs',
checkpoint_name, 'predictions',
results_dir, 'kitti', checkpoint_number,
uncertainty_method, 'data')
else:
prediction_dir = os.path.join(core.data_dir(), 'outputs',
checkpoint_name, 'predictions',
results_dir, checkpoint_number, 'data')
frames_list = os.listdir(prediction_dir)
index = random.randint(0, len(frames_list) - 1)
frame_id = int(frames_list[index][0:6])
# frame_id = 27 # Out of distribution example
frame_id = 4079
# frame_id = 169 # Many Cars, Hard
# frame_id = 2290 # Many Cars, Occlusions
# frame_id = 1941 # Many Cars, Horizontal Direction
# frame_id = 4032 # Orientation
# frame_id = 104 # Weird Orientation
# frame_id = 7047 # Weird Orientation
# frame_id = 6632 # Very hard orientations
# frame_id = 195 # Single Pedestrian
# frame_id = 1574 # Single Pedestrian
# frame_id = 332 # Multiple Hard Pedestrians
# frame_id = 1193 # Multiple Hard Pedestrians
# frame_id = 1274 # Multiple Cyclists
print('Showing Frame: %d' % frame_id)
#############
# Read Frame
#############
im_path = image_dir + '/{:06d}.png'.format(frame_id)
image = cv2.imread(im_path)
label_path = label_dir + '/{:06d}.txt'.format(frame_id)
gt_classes_hard, gt_boxes_hard = read_labels(label_path)
prediction_path = prediction_dir + '/{:06d}.txt'.format(frame_id)
prediction_classes, prediction_boxes, prediction_scores = read_predictions(
prediction_path)
max_ious = np.zeros(prediction_boxes.shape[0])
# Compute IOU between each prediction and the ground truth boxes
if gt_boxes_hard.size > 0 and prediction_boxes.size > 0:
for obj_idx in range(prediction_boxes.shape[0]):
obj_iou_fmt = prediction_boxes[obj_idx]
ious_2d = two_d_iou(obj_iou_fmt, gt_boxes_hard)
max_iou = np.amax(ious_2d)
max_ious[obj_idx] = max_iou
#########################################################
# Draw GT and Prediction Boxes
#########################################################
# Transform Predictions to left and right images
image_out = draw_box_2d(image,
gt_boxes_hard,
gt_classes_hard,
line_width=2,
dataset='kitti',
is_gt=True)
image_out = draw_box_2d(image_out,
prediction_boxes,
prediction_classes,
line_width=2,
is_gt=False,
dataset='kitti',
text_to_plot=max_ious,
plot_text=True)
if results_dir == 'testing':
cv2.imshow('Detections from ' + uncertainty_method, image_out)
else:
cv2.imshow('Validation Set Detections', image_out)
cv2.waitKey()
def main():
# set root_dir to the correct path to your dataset folder
root_dir = core.data_dir(
) + '/datasets/argoverse/argoverse-tracking/sample/'
vtk_window_size = (1280, 720)
argoverse_loader = ArgoverseTrackingLoader(root_dir)
print('Total number of logs:', len(argoverse_loader))
argoverse_loader.print_all()
for i, argoverse_data in enumerate(argoverse_loader):
if i >= 3:
break
print(argoverse_data)
argoverse_data = argoverse_loader[0]
print(argoverse_data)
argoverse_data = argoverse_loader.get(
'c6911883-1843-3727-8eaa-41dc8cda8993')
print(argoverse_data)
log_id = 'c6911883-1843-3727-8eaa-41dc8cda8993' # argoverse_loader.log_list[55]
frame_idx = 150
camera = argoverse_loader.CAMERA_LIST[0]
argoverse_data = argoverse_loader.get(log_id)
city_name = argoverse_data.city_name
print('-------------------------------------------------------')
print(f'Log: {log_id}, \n\tframe: {frame_idx}, camera: {camera}')
print('-------------------------------------------------------\n')
lidar_points = argoverse_data.get_lidar(frame_idx)
img = argoverse_data.get_image_sync(frame_idx, camera=camera)
objects = argoverse_data.get_label_object(frame_idx)
calib = argoverse_data.get_calibration(camera)
# TODO: Calculate point colours
vtk_pc = VtkPointCloudGlyph()
vtk_pc.set_points(lidar_points)
vtk_pc.set_point_size(2)
vtk_renderer = demo_utils.setup_vtk_renderer()
vtk_render_window = demo_utils.setup_vtk_render_window(
'Argoverse Demo', vtk_window_size, vtk_renderer)
vtk_axes = vtk.vtkAxesActor()
vtk_axes.SetTotalLength(2, 2, 2)
vtk_renderer.AddActor(vtk_axes)
vtk_renderer.AddActor(vtk_pc.vtk_actor)
current_cam = vtk_renderer.GetActiveCamera()
current_cam.SetViewUp(0, 0, 1)
current_cam.SetPosition(-50, 0, 15)
current_cam.SetFocalPoint(30, 0, 0)
vtk_interactor = demo_utils.setup_vtk_interactor(vtk_render_window)
vtk_interactor.Start()
def main():
"""Interpolates the lidar point cloud using IP-Basic
and saves a dense depth map of the scene.
https://github.com/kujason/ip_basic
"""
##############################
# Options
##############################
kitti_raw_dir = os.path.expanduser(
'/local-scratch/hadi/datasets/kitti_raw')
drive_id = '2011_09_26_drive_0001_sync'
# drive_id = '2011_09_26_drive_0039_sync'
raw_data = raw_utils.get_raw_data(drive_id, kitti_raw_dir)
# Fill algorithm ('ip_basic_{...}')
fill_type = 'multiscale'
save_depth_maps = True
out_depth_map_dir = core.data_dir() + '/results/{}/depth_02_{}'.format(
drive_id, fill_type)
##############################
# End of Options
##############################
os.makedirs(out_depth_map_dir, exist_ok=True)
# Rolling average array of times for time estimation
avg_time_arr_length = 5
last_fill_times = np.repeat([1.0], avg_time_arr_length)
last_total_times = np.repeat([1.0], avg_time_arr_length)
cam_p = raw_data.calib.P_rect_20
image_shape = raw_data.get_cam2(0).size[::-1]
frames_to_use = raw_data.velo_files
num_frames = len(frames_to_use)
for frame_idx, velo_path in enumerate(frames_to_use):
# Calculate average time with last n fill times
avg_fill_time = np.mean(last_fill_times)
avg_total_time = np.mean(last_total_times)
# Print progress
sys.stdout.write('\rProcessing {} / {}, Avg Fill Time: {:.5f}s, '
'Avg Time: {:.5f}s, Est Time: {:.3f}s'.format(
frame_idx, num_frames - 1, avg_fill_time,
avg_total_time,
avg_total_time * (num_frames - frame_idx)))
sys.stdout.flush()
# Start timing
start_total_time = time.time()
# Load point cloud
velo_points = raw_data.get_velo(frame_idx)[:, 0:3]
velo_pc_padded = transform_utils.pad_pc(velo_points.T)
cam0_point_cloud = raw_data.calib.R_rect_00 @ raw_data.calib.T_cam0_velo @ velo_pc_padded
cam0_point_cloud, _ = obj_utils.filter_pc_to_area(
cam0_point_cloud[0:3],
area_extents=np.asarray([[-40, 40], [-3, 5], [0, 80]]))
# Project point cloud to create depth map
projected_depths = depth_map_utils.project_depths(
cam0_point_cloud, cam_p, image_shape)
# Fill depth map
if fill_type == 'multiscale':
start_fill_time = time.time()
final_depth_map, _ = ip_basic.fill_in_multiscale(projected_depths)
end_fill_time = time.time()
else:
raise ValueError('Invalid fill algorithm')
# Save depth maps
if save_depth_maps:
out_depth_map_path = out_depth_map_dir + '/{:010d}.png'.format(
frame_idx)
depth_map_utils.save_depth_map(out_depth_map_path, final_depth_map)
# Stop timing
end_total_time = time.time()
# Update fill times
last_fill_times = np.roll(last_fill_times, -1)
last_fill_times[-1] = end_fill_time - start_fill_time
# Update total times
last_total_times = np.roll(last_total_times, -1)
last_total_times[-1] = end_total_time - start_total_time
def main():
# Load configuration
with open("config/default.yaml") as file:
config = yaml.full_load(file)
# Get directories
detections_dir = os.path.join(core.data_dir(), "detections")
dataset_dir = os.path.join("~/Kitti/tracking")
results_dir = os.path.join(core.top_dir(), "results")
plot_dir = os.path.join(results_dir, "plots")
make_dir(plot_dir)
ylabels = [
r"$x$ [m]", r"$y$ [m]", r"$z$ [m]", r"$\theta$ [rad]", r"$l$ [m]",
r"$h$ [m]", r"$w$ [m]"
]
for split in config["splits"]:
for class_ in config["classes"]:
seq_id = 15
gt_track_id = 2
# Load ground truth
sequence = KittiSequence(detections_dir=detections_dir,
dataset_dir=dataset_dir,
seq_id=seq_id,
split=split,
class_=class_)
gt_track = sequence.get_track(gt_track_id)
gt_track = gt_track[:, 2:]
# Load track
track_dir = os.path.join(results_dir, "tracks", split, class_,
str(seq_id).zfill(4))
track_id = "38882006386010313624536118754.npy"
x_file = os.path.join(track_dir, "x", track_id)
P_file = os.path.join(track_dir, "P", track_id)
x = np.load(x_file)
x = x[:7, :]
P = np.load(P_file)
P = P[:7, :7, :]
# Compute error and std dev
t = np.arange(x.shape[1])
error = x - gt_track
error[3, :] = np.unwrap(error[3, :])
var = np.transpose(P.diagonal())
std_dev = np.sqrt(var)
fig = plt.figure(figsize=(12, 7))
plot_file = os.path.join(plot_dir, "error.png")
for i, ylabel in enumerate(ylabels):
ax = fig.add_subplot(241 + i)
line1 = ax.plot(t, error[i, :])
line2 = ax.plot(t, 3 * std_dev[i, :], 'r--', t,
-3 * std_dev[i, :], 'r--')[0]
ax.set_xlabel('Frame')
ax.set_ylabel(ylabel)
fig.legend([line1, line2],
labels=['Error', 'Uncertainty Envelope'],
bbox_to_anchor=(0.97, 0.28),
loc="center right")
plt.tight_layout()
plt.savefig(plot_file)
plt.close()
def main():
# Load configuration
with open("config/default.yaml") as file:
config = yaml.full_load(file)
# Get directories
detections_dir = os.path.join(core.data_dir(), "detections")
dataset_dir = os.path.join("~/Kitti/tracking")
results_dir = os.path.join(core.top_dir(), "results")
duration = 0
frames = 0
for split in config["splits"]:
for class_ in config["classes"]:
for seq_id in range(config["num_sequences"]):
print("Processing: {}".format(seq_id))
# Load detections
sequence_detections = KittiSequence(
detections_dir=detections_dir,
dataset_dir=dataset_dir,
seq_id=seq_id,
split=split,
class_=class_)
start = time.time()
sequence_tracker = Tracker(sequence_detections)
sequence_tracker.run()
duration = duration + (time.time() - start)
frames = frames + len(sequence_detections)
# Output text predictions
if config["output_text"]:
text_dir = os.path.join(results_dir, "text", split, class_)
# Make file if doesn't exists
if not os.path.exists(text_dir):
os.makedirs(text_dir)
text_file = os.path.join(text_dir,
str(seq_id).zfill(4) + ".txt")
sequence_tracker.generate_text_output(
output_file=text_file)
# Output bounding box visualization
if config["output_vis"]:
vis_path = os.path.join(results_dir, "vis", split, class_,
str(seq_id).zfill(4))
sequence_tracker.generate_visualization(
output_path=vis_path)
# Output error and covariance plots
if config["output_tracks"]:
track_dir = os.path.join(results_dir, "tracks", split,
class_,
str(seq_id).zfill(4))
# Make file if doesn't exists
if not os.path.exists(track_dir):
os.makedirs(track_dir)
sequence_tracker.generate_track_output(
output_path=track_dir)
print("FPS {}".format(frames / duration))
def main():
"""Comparison of ground truth and ORBSLAM2 poses
https://github.com/raulmur/ORB_SLAM2
"""
####################
# Options
####################
odom_dataset_dir = os.path.expanduser('~/Kitti/odometry/dataset')
# sequence = '02'
# sequence = '03'
# sequence = '08'
sequence = '09'
# sequence = '10'
# sequence = '11'
# sequence = '12'
# vtk_window_size = (1280, 720)
vtk_window_size = (960, 540)
# vtk_window_size = (400, 300)
point_cloud_source = 'lidar'
# point_cloud_source = 'fast'
# point_cloud_source = 'multiscale'
# first_frame_idx = None
# first_frame_pose = None
# Setup odometry dataset handler
odom_dataset = pykitti.odometry(odom_dataset_dir, sequence)
# # Check that velo length matches timestamps?
# if len(odom_dataset.velo_files) != len(odom_dataset.timestamps):
# raise ValueError('velo files and timestamps have different length!')
frame_range = (0, len(odom_dataset.timestamps))
# frame_range = (0, 100)
# camera_viewpoint = 'front'
camera_viewpoint = 'elevated'
# camera_viewpoint = 'bev'
buffer_size = 50
buffer_update = 10
save_screenshots = False
##############################
vtk_renderer = demo_utils.setup_vtk_renderer()
vtk_render_window = demo_utils.setup_vtk_render_window(
'Overlaid Point Cloud', vtk_window_size, vtk_renderer)
if save_screenshots:
vtk_win_to_img_filter, vtk_png_writer = vtk_utils.setup_screenshots(
vtk_render_window)
vtk_interactor = vtk.vtkRenderWindowInteractor()
vtk_interactor.SetRenderWindow(vtk_render_window)
vtk_interactor.SetInteractorStyle(
vtk_utils.ToggleActorsInteractorStyle(None, vtk_renderer))
vtk_interactor.Initialize()
cam_p2 = odom_dataset.calib.P_rect_20
# Load poses
cam0_ref_poses_orbslam = np.loadtxt(
odom_dataset_dir + '/poses_orbslam2/{}.txt'.format(sequence))
cam0_ref_poses_orbslam = np.pad(cam0_ref_poses_orbslam, ((0, 0), (0, 4)),
mode='constant')
cam0_ref_poses_orbslam[:, -1] = 1
cam0_ref_poses_orbslam = cam0_ref_poses_orbslam.reshape((-1, 4, 4))
cam0_ref_poses_gt = np.asarray(odom_dataset.poses, np.float32)
# Setup camera
if camera_viewpoint == 'front':
cam0_curr_vtk_cam_pos = [0.0, 0.0, 0.0, 1.0]
cam0_curr_vtk_focal_point = [0.0, 0.0, 20.0, 1.0]
elif camera_viewpoint == 'elevated':
cam0_curr_vtk_cam_pos = [0.0, -5.0, -15.0, 1.0]
cam0_curr_vtk_focal_point = [0.0, 0.0, 20.0, 1.0]
elif camera_viewpoint == 'bev':
# camera_zoom = 1.0
cam0_curr_vtk_cam_pos = [0.0, -50.0, 10.0, 1.0]
# camera_pos = (0.0, 0.0, 0.0)
cam0_curr_vtk_focal_point = [0.0, 0.0, 15.0, 1.0]
else:
raise ValueError('Invalid camera_pos', camera_viewpoint)
# Create VtkAxes
vtk_axes = vtk.vtkAxesActor()
vtk_axes.SetTotalLength(2, 2, 2)
vtk_renderer.AddActor(vtk_axes)
vtk_pc_poses_orbslam = VtkPointCloudGlyph()
vtk_pc_poses_gt = VtkPointCloudGlyph()
vtk_pc_poses_orbslam.vtk_actor.GetProperty().SetPointSize(5)
vtk_pc_poses_gt.vtk_actor.GetProperty().SetPointSize(5)
actor_buffer = []
for frame_idx in range(*frame_range):
print('{} / {}'.format(frame_idx, len(odom_dataset.timestamps) - 1))
# Load next frame data
load_start_time = time.time()
rgb_load_start = time.time()
bgr_image = cv2.imread(odom_dataset.cam2_files[frame_idx])
print('rgb_load', time.time() - rgb_load_start)
if point_cloud_source == 'lidar':
velo_curr_points_padded = get_velo_points(odom_dataset, frame_idx)
# Velo in cam0_curr
cam0_curr_pc_all_padded = odom_dataset.calib.T_cam0_velo @ velo_curr_points_padded.T
pass
# cam0_curr_points = cam0_curr_pc.T
elif point_cloud_source in ['fast', 'multiscale']:
# depth_map_path = depth_map_dir + '/{:010d}.png'.format(frame_idx)
# depth_map = depth_map_utils.read_depth_map(depth_map_path)
# points_cam2 = depth_map_utils.get_depth_point_cloud(depth_map, cam_p).T
raise NotImplementedError()
else:
raise ValueError('Invalid point cloud source')
print('load\t\t', time.time() - load_start_time)
# Project velodyne points
projection_start_time = time.time()
if point_cloud_source == 'lidar':
# Project into image2
points_in_img2 = calib_utils.project_pc_to_image2(
cam0_curr_pc_all_padded, cam_p2)
points_in_img2_int = np.round(points_in_img2).astype(np.int32)
image_filter = obj_utils.points_in_img_filter(
points_in_img2_int, bgr_image.shape)
cam0_curr_pc_padded = cam0_curr_pc_all_padded[:, image_filter]
# points = points_cam2[image_mask]
points_in_img_int_valid = points_in_img2_int[:, image_filter]
point_colours = bgr_image[points_in_img_int_valid[1],
points_in_img_int_valid[0]]
else:
raise ValueError('Invalid point_cloud_source', point_cloud_source)
print('projection\t', time.time() - projection_start_time)
# Get pose
cam0_ref_pose_orbslam = cam0_ref_poses_orbslam[frame_idx]
cam0_ref_pose_gt = cam0_ref_poses_gt[frame_idx]
tf_cam0_ref_cam0_curr = cam0_ref_pose_orbslam
cam0_ref_pc_padded = tf_cam0_ref_cam0_curr @ cam0_curr_pc_padded
# VtkPointCloud
vtk_pc = VtkPointCloudGlyph()
vtk_pc.vtk_actor.GetProperty().SetPointSize(2)
vtk_pc_start_time = time.time()
vtk_pc.set_points(cam0_ref_pc_padded[0:3].T, point_colours)
print('vtk_pc\t\t', time.time() - vtk_pc_start_time)
# Display orbslam pose
vtk_pc_poses_orbslam.set_points(
cam0_ref_poses_orbslam[max(0, frame_idx - buffer_size):frame_idx +
1, 0:3, 3],
np.tile([255, 0, 0], [buffer_size, 1]))
# Display gt pose
vtk_pc_poses_gt.vtk_actor.GetProperty().SetPointSize(5)
vtk_pc_poses_gt.set_points(
cam0_ref_poses_gt[max(0, frame_idx - buffer_size):frame_idx + 1,
0:3, 3], np.tile([0, 255, 0], [buffer_size, 1]))
# Add vtk actors
vtk_renderer.AddActor(vtk_pc.vtk_actor)
vtk_renderer.AddActor(vtk_pc_poses_orbslam.vtk_actor)
vtk_renderer.AddActor(vtk_pc_poses_gt.vtk_actor)
cam0_ref_vtk_cam_pos = tf_cam0_ref_cam0_curr.dot(cam0_curr_vtk_cam_pos)
cam0_ref_vtk_focal_point = tf_cam0_ref_cam0_curr.dot(
cam0_curr_vtk_focal_point)
current_cam = vtk_renderer.GetActiveCamera()
vtk_renderer.ResetCamera()
current_cam.SetViewUp(0, -1, 0)
current_cam.SetPosition(cam0_ref_vtk_cam_pos[0:3])
current_cam.SetFocalPoint(*cam0_ref_vtk_focal_point[0:3])
current_cam.Zoom(0.5)
vtk_renderer.ResetCameraClippingRange()
# Render
render_start_time = time.time()
vtk_render_window.Render()
print('render\t\t', time.time() - render_start_time)
actor_buffer.append(vtk_pc.vtk_actor)
if len(actor_buffer) > buffer_size:
if frame_idx % buffer_update != 0:
actor_buffer[frame_idx - buffer_size].SetVisibility(0)
if save_screenshots:
screenshot_start_time = time.time()
screenshot_path = core.data_dir() + '/temp/{:010d}.png'.format(
frame_idx)
vtk_utils.save_screenshot(screenshot_path, vtk_win_to_img_filter,
vtk_png_writer)
print('screenshot\t', time.time() - screenshot_start_time)
print('---')
for vtk_actor in actor_buffer:
vtk_actor.SetVisibility(1)
print('Done')
# Keep window open
vtk_interactor.Start()
def main():
#########################################################
# Specify Source Folders and Parameters For Frame Reader
#########################################################
data_split_dir = 'val'
# Specify whether the validation or inference results need to be
# visualized.
# results_dir = 'validation' # Or testing
results_dir = 'testing'
# sample_free, anchor_redundancy, black_box, bayes_od_none,
# bayes_od_ci_fast. bayes_od_ci,or bayes_od_ici
uncertainty_method = 'bayes_od_ci_fast'
dataset_dir = os.path.expanduser('~/Datasets/bdd100k')
image_dir = os.path.join(dataset_dir, 'images', '100k', data_split_dir)
label_file_name = os.path.join(dataset_dir, 'labels',
data_split_dir) + '.json'
checkpoint_name = 'retinanet_bdd'
checkpoint_number = '101'
if results_dir == 'testing':
prediction_dir = os.path.join(core.data_dir(), 'outputs',
checkpoint_name, 'predictions',
results_dir, 'bdd', checkpoint_number,
uncertainty_method, 'data')
else:
prediction_dir = os.path.join(core.data_dir(), 'outputs',
checkpoint_name, 'predictions',
results_dir, checkpoint_number, 'data')
prediction_file_name = os.path.join(prediction_dir, 'predictions.json')
frames_list = os.listdir(image_dir)
index = random.randint(0, len(frames_list))
frame_id = frames_list[index]
print('Showing Frame ID:' + frame_id)
#############
# Read Frame
#############
im_path = image_dir + '/' + frame_id
image = cv2.imread(im_path)
all_labels = json.load(open(label_file_name, 'r'))
category_gt, boxes_2d_gt, _ = read_bdd_format(
frame_id,
all_labels,
categories=['car', 'truck', 'bus', 'person', 'rider', 'bike', 'motor'])
all_predictions = json.load(open(prediction_file_name, 'r'))
category_pred, boxes_2d_pred, _ = read_bdd_format(
frame_id,
all_predictions,
categories=['car', 'truck', 'bus', 'person', 'rider', 'bike', 'motor'])
max_ious = np.zeros(boxes_2d_pred.shape[0])
# Compute IOU between each prediction and the ground truth boxes
if boxes_2d_gt.size > 0 and boxes_2d_pred.size > 0:
for obj_idx in range(boxes_2d_pred.shape[0]):
obj_iou_fmt = boxes_2d_pred[obj_idx]
ious_2d = two_d_iou(obj_iou_fmt, boxes_2d_gt)
max_iou = np.amax(ious_2d)
max_ious[obj_idx] = max_iou
#########################################################
# Draw GT and Prediction Boxes
#########################################################
# Transform Predictions to left and right images
image_out = draw_box_2d(image,
boxes_2d_gt,
category_gt,
line_width=2,
dataset='bdd',
is_gt=True)
image_out = draw_box_2d(image_out,
boxes_2d_pred,
category_pred,
line_width=2,
is_gt=False,
dataset='bdd',
text_to_plot=max_ious,
plot_text=True)
if results_dir == 'testing':
cv2.imshow('Detections from ' + uncertainty_method, image_out)
else:
cv2.imshow('Validation Set Detections', image_out)
cv2.waitKey()
def main():
##############################
# Options
##############################
"""Note: Run scripts/depth_completion/save_depth_maps_raw.py first.
Demo to show depth completed point clouds
"""
raw_dir = os.path.expanduser('~/Kitti/raw')
drive_id = '2011_09_26_drive_0023_sync'
vtk_window_size = (1280, 720)
max_fps = 30.0
# point_cloud_source = 'lidar'
# fill_type = None
point_cloud_source = 'depth'
fill_type = 'multiscale'
# Load raw data
drive_date = drive_id[0:10]
drive_num_str = drive_id[17:21]
raw_data = pykitti.raw(raw_dir, drive_date, drive_num_str)
# Check that velo length matches timestamps?
if len(raw_data.velo_files) != len(raw_data.timestamps):
raise ValueError('velo files and timestamps have different length!')
frame_range = (0, len(raw_data.timestamps))
##############################
min_loop_time = 1.0 / max_fps
vtk_renderer = demo_utils.setup_vtk_renderer()
vtk_render_window = demo_utils.setup_vtk_render_window(
'Overlaid Point Cloud', vtk_window_size, vtk_renderer)
# Setup camera
current_cam = vtk_renderer.GetActiveCamera()
current_cam.SetViewUp(0, 0, 1)
current_cam.SetPosition(0.0, -8.0, -15.0)
current_cam.SetFocalPoint(0.0, 0.0, 20.0)
current_cam.Zoom(0.7)
# Create VtkAxes
vtk_axes = vtk.vtkAxesActor()
vtk_axes.SetTotalLength(2, 2, 2)
# Setup interactor
vtk_interactor = vtk.vtkRenderWindowInteractor()
vtk_interactor.SetRenderWindow(vtk_render_window)
vtk_interactor.SetInteractorStyle(
vtk_utils.ToggleActorsInteractorStyle(None, vtk_renderer, current_cam,
vtk_axes))
vtk_interactor.Initialize()
if point_cloud_source not in ['lidar', 'depth']:
raise ValueError(
'Invalid point cloud source {}'.format(point_cloud_source))
cam_p = raw_data.calib.P_rect_20
# Point cloud
vtk_pc = VtkPointCloudGlyph()
# Add actors
vtk_renderer.AddActor(vtk_axes)
vtk_renderer.AddActor(vtk_pc.vtk_actor)
for frame_idx in range(*frame_range):
loop_start_time = time.time()
print('{} / {}'.format(frame_idx, len(raw_data.timestamps) - 1))
# Load next frame data
load_start_time = time.time()
rgb_image = np.asarray(raw_data.get_cam2(frame_idx))
bgr_image = rgb_image[..., ::-1]
if point_cloud_source == 'lidar':
velo_points = get_velo_points(raw_data, frame_idx)
# Transform point cloud to cam_0 frame
velo_curr_points_padded = np.pad(velo_points, [[0, 0], [0, 1]],
constant_values=1.0,
mode='constant')
cam0_curr_pc_all_padded = raw_data.calib.T_cam0_velo @ velo_curr_points_padded.T
# Project velodyne points
projection_start_time = time.time()
points_in_img = calib_utils.project_pc_to_image(
cam0_curr_pc_all_padded[0:3], cam_p)
points_in_img_int = np.round(points_in_img).astype(np.int32)
print('projection\t', time.time() - projection_start_time)
image_filter = obj_utils.points_in_img_filter(
points_in_img_int, bgr_image.shape)
cam0_curr_pc = cam0_curr_pc_all_padded[0:3, image_filter]
points_in_img_int_valid = points_in_img_int[:, image_filter]
point_colours = bgr_image[points_in_img_int_valid[1],
points_in_img_int_valid[0]]
elif point_cloud_source == 'depth':
depth_maps_dir = core.data_dir() + \
'/depth_completion/raw/{}/depth_02_{}'.format(drive_id, fill_type)
depth_map_path = depth_maps_dir + '/{:010d}.png'.format(frame_idx)
depth_map = depth_map_utils.read_depth_map(depth_map_path)
cam0_curr_pc = depth_map_utils.get_depth_point_cloud(
depth_map, cam_p)
point_colours = bgr_image.reshape(-1, 3)
# Mask to valid points
valid_mask = (cam0_curr_pc[2] != 0)
cam0_curr_pc = cam0_curr_pc[:, valid_mask]
point_colours = point_colours[valid_mask]
else:
raise ValueError('Invalid point cloud source')
print('load\t\t', time.time() - load_start_time)
# VtkPointCloud
vtk_pc_start_time = time.time()
vtk_pc.set_points(cam0_curr_pc.T, point_colours)
print('vtk_pc\t\t', time.time() - vtk_pc_start_time)
# Reset the clipping range to show all points
vtk_renderer.ResetCameraClippingRange()
# Render
render_start_time = time.time()
vtk_render_window.Render()
print('render\t\t', time.time() - render_start_time)
# Pause to keep frame rate under max
loop_run_time = time.time() - loop_start_time
print('loop\t\t', loop_run_time)
if loop_run_time < min_loop_time:
time.sleep(min_loop_time - loop_run_time)
print('---')
print('Done')
# Keep window open
vtk_interactor.Start()
def setup_config(args, random_seed=None, is_testing=False):
"""
Sets up config node with probabilistic detectron elements. Also sets up a fixed random seed for all scientific
computing libraries, and sets up all supported datasets as instances of coco.
Args:
args (Namespace): args from argument parser
random_seed (int): set a fixed random seed throughout torch, numpy, and python
is_testing (bool): set to true if inference. If true function will return an error if checkpoint directory not
already existing.
Returns:
(CfgNode) detectron2 config object
"""
# Get default detectron config file
cfg = get_cfg()
add_detr_config(cfg)
add_probabilistic_config(cfg)
# Update default config file with custom config file
configs_dir = core.configs_dir()
args.config_file = os.path.join(configs_dir, args.config_file)
cfg.merge_from_file(args.config_file)
# Add dropout rate for faster RCNN box head
cfg.MODEL.ROI_BOX_HEAD.DROPOUT_RATE = cfg.MODEL.PROBABILISTIC_MODELING.DROPOUT_RATE
# Update config with inference configurations. Only applicable for when in
# probabilistic inference mode.
if args.inference_config != "":
args.inference_config = os.path.join(configs_dir,
args.inference_config)
cfg.merge_from_file(args.inference_config)
# Create output directory
model_name = os.path.split(os.path.split(args.config_file)[0])[-1]
dataset_name = os.path.split(
os.path.split(os.path.split(args.config_file)[0])[0])[-1]
cfg['OUTPUT_DIR'] = os.path.join(core.data_dir(), dataset_name, model_name,
os.path.split(args.config_file)[-1][:-5],
'random_seed_' + str(random_seed))
if is_testing:
if not os.path.isdir(cfg['OUTPUT_DIR']):
raise NotADirectoryError(
"Checkpoint directory {} does not exist.".format(
cfg['OUTPUT_DIR']))
os.makedirs(cfg['OUTPUT_DIR'], exist_ok=True)
# copy config file to output directory
copyfile(
args.config_file,
os.path.join(cfg['OUTPUT_DIR'],
os.path.split(args.config_file)[-1]))
# Freeze config file
cfg['SEED'] = random_seed
cfg.freeze()
# Initiate default setup
default_setup(cfg, args)
# Setup logger for probabilistic detectron module
setup_logger(output=cfg.OUTPUT_DIR,
distributed_rank=comm.get_rank(),
name="Probabilistic Detectron")
# Set a fixed random seed for all numerical libraries
if random_seed is not None:
torch.manual_seed(random_seed)
np.random.seed(random_seed)
random.seed(random_seed)
# Setup datasets
if args.image_corruption_level != 0:
image_root_corruption_prefix = '_' + str(args.image_corruption_level)
else:
image_root_corruption_prefix = None
dataset_dir = os.path.expanduser(args.dataset_dir)
# Handle cases when this function has been called multiple times. In that case skip fully.
# Todo this is very bad practice, should fix.
try:
setup_all_datasets(
dataset_dir,
image_root_corruption_prefix=image_root_corruption_prefix)
return cfg
except AssertionError:
return cfg
