def __getitem__(self, index):
example = self.examples[index]
img_id = example["img_id"]
calib_path = self.calib_dir + img_id + ".txt"
calib = calibread(calib_path)
camera_matrix = calib['P2']
u_min = example["u_min"] # (left)
u_max = example["u_max"] # (rigth)
v_min = example["v_min"] # (top)
v_max = example["v_max"] # (bottom)
score_2d = example["score_2d"]
input_2Dbbox = np.array([u_min, u_max, v_min, v_max])
w = u_max - u_min
h = v_max - v_min
u_center = u_min + w / 2.0
v_center = v_min + h / 2.0
########################################################################
# get the input 2dbbox img crop and resize to 224 x 224:
########################################################################
img_path = self.img_dir + img_id + ".png"
img = cv2.imread(img_path, -1)
bbox_2d_img = img[int(np.max([0, v_min])):int(v_max),
int(np.max([0, u_min])):int(u_max)]
bbox_2d_img = cv2.resize(bbox_2d_img, (224, 224))
# # # # # debug visualization:
#cv2.imshow("test", bbox_2d_img)
#cv2.waitKey(0)
# # # # #
########################################################################
# normalize the 2dbbox img crop:
########################################################################
bbox_2d_img = bbox_2d_img / 255.0
bbox_2d_img = bbox_2d_img - np.array([0.485, 0.456, 0.406])
bbox_2d_img = bbox_2d_img / np.array([0.229, 0.224, 0.225
]) # (shape: (H, W, 3))
bbox_2d_img = np.transpose(bbox_2d_img,
(2, 0, 1)) # (shape: (3, H, W))
bbox_2d_img = bbox_2d_img.astype(np.float32)
########################################################################
# convert numpy -> torch:
########################################################################
bbox_2d_img = torch.from_numpy(
bbox_2d_img) # (shape: (3, H, W) = (3, 224, 224))
return (bbox_2d_img, img_id, self.mean_car_size, w, h, u_center,
v_center, input_2Dbbox, camera_matrix, self.mean_distance,
score_2d)
lineType=cv2.LINE_AA,
thickness=2)
return img
for sequence in ["0001", "0002", "0007", "0011"]:
print(sequence)
project_dir = "/home/fregu856/3DOD_thesis/" # NOTE! you'll have to adapt this for your file structure
data_dir = project_dir + "data/kitti/tracking/testing/"
img_dir = data_dir + "image_02/" + sequence + "/"
calib_path = project_dir + "data/kitti/meta/tracking/testing/calib/" + sequence + ".txt" # NOTE! kitti/meta
lidar_dir = data_dir + "velodyne/" + sequence + "/"
calib = calibread(calib_path)
P2 = calib["P2"]
Tr_velo_to_cam_orig = calib["Tr_velo_to_cam"]
R0_rect_orig = calib["R0_rect"]
R0_rect = np.eye(4)
R0_rect[0:3, 0:3] = R0_rect_orig
Tr_velo_to_cam = np.eye(4)
Tr_velo_to_cam[0:3, :] = Tr_velo_to_cam_orig
# NOTE! here you can choose what model's output you want to visualize
# Frustum-PointNet:
with open(
"/home/fregu856/3DOD_thesis/training_logs/model_Frustum-PointNet_eval_test_seq/eval_dict_test_seq_%s.pkl"
% sequence, "rb"
"000000", "000001", "000002", "000003", "000004", "000005",
"000006", "000007", "000008", "000009", "000010"
]:
print img_id
bbox_dicts = eval_dict[img_id]
img = cv2.imread(img_dir + img_id + ".png", -1)
lidar_path = lidar_dir + img_id + ".bin"
point_cloud = np.fromfile(lidar_path, dtype=np.float32).reshape(-1, 4)
# remove points that are located behind the camera:
point_cloud = point_cloud[point_cloud[:, 0] > -2.5, :]
calib = calibread(calib_dir + img_id + ".txt")
P2 = calib["P2"]
Tr_velo_to_cam_orig = calib["Tr_velo_to_cam"]
R0_rect_orig = calib["R0_rect"]
#
R0_rect = np.eye(4)
R0_rect[0:3, 0:3] = R0_rect_orig
#
Tr_velo_to_cam = np.eye(4)
Tr_velo_to_cam[0:3, :] = Tr_velo_to_cam_orig
point_cloud_xyz = point_cloud[:, 0:3]
point_cloud_xyz_hom = np.ones((point_cloud.shape[0], 4))
point_cloud_xyz_hom[:, 0:
3] = point_cloud[:, 0:
3] # (point_cloud_xyz_hom has shape (num_points, 4))
