def __init__(self, root, split=None):
self.data = LyftDataset(data_path=root,
json_path=os.path.join(
root,
os.path.basename(os.path.normpath(root))),
verbose=False)
self.explorer = LyftDatasetExplorer(self.data)
super().__init__(root, split)
def render_ann(ann_token, dataLyft3D):
ann_record = dataLyft3D.get("sample_annotation", ann_token)
sample_record = dataLyft3D.get("sample", ann_record["sample_token"])
fig, axes = plt.subplots(1, 1, figsize=(12, 6))
# Figure out which camera the object is fully visible in (this may return nothing)
boxes, cam = [], []
cams = [key for key in sample_record["data"].keys() if "CAM" in key]
for cam in cams:
_, boxes, _ = dataLyft3D.get_sample_data(
sample_record["data"][cam],
box_vis_level=BoxVisibility.ANY,
selected_anntokens=[ann_token])
if len(boxes) > 0:
break # We found an image that matches. Let's abort.
cam = sample_record["data"][cam]
data_path, boxes, camera_intrinsic = dataLyft3D.get_sample_data(
cam, selected_anntokens=[ann_token])
im = Image.open(data_path)
axes.imshow(im)
axes.set_title(dataLyft3D.get("sample_data", cam)["channel"])
axes.axis("off")
axes.set_aspect("equal")
for box in boxes:
c = np.array(LyftDatasetExplorer.get_color(box.name)) / 255.0
render_box(box,
axes,
view=camera_intrinsic,
normalize=True,
colors=(c, c, c),
im=im)
def __init__(self,
data_path: str,
json_path: str,
verbose: bool = True,
map_resolution: float = 0.1):
"""Loads database and creates reverse indexes and shortcuts.
Args:
data_path: Path to the tables and data.
json_path: Path to the folder with json files
verbose: Whether to print status messages during load.
map_resolution: Resolution of maps (meters).
"""
self.data_path = Path(data_path).expanduser().absolute()
self.json_path = Path(json_path)
self.table_names = [
"category",
"attribute",
"sensor",
"calibrated_sensor",
"ego_pose",
"log",
"scene",
"sample",
"sample_data",
"map",
]
start_time = time.time()
# Explicitly assign tables to help the IDE determine valid class members.
self.category = self.__load_table__("category")
self.attribute = self.__load_table__("attribute")
self.sensor = self.__load_table__("sensor")
self.calibrated_sensor = self.__load_table__("calibrated_sensor")
self.ego_pose = self.__load_table__("ego_pose")
self.log = self.__load_table__("log")
self.scene = self.__load_table__("scene")
self.sample = self.__load_table__("sample")
self.sample_data = self.__load_table__("sample_data")
self.map = self.__load_table__("map")
if verbose:
for table in self.table_names:
print("{} {},".format(len(getattr(self, table)), table))
print(
"Done loading in {:.1f} seconds.\n======".format(time.time() -
start_time))
# Initialize LyftDatasetExplorer class
self.explorer = LyftDatasetExplorer(self)
# Make reverse indexes for common lookups.
self.__make_reverse_index__(verbose)
def get_sample_images(sample_token: str, ax=None, lyftd=level5data):
record = lyftd.get("sample", sample_token)
# Separate RADAR from LIDAR and vision.
radar_data = {}
nonradar_data = {}
for channel, token in record["data"].items():
sd_record = lyftd.get("sample_data", token)
sensor_modality = sd_record["sensor_modality"]
if sensor_modality in ["lidar", "camera"]:
nonradar_data[channel] = token
else:
radar_data[channel] = token
# get projective matrix
for channel, token in nonradar_data.items():
sd_record = lyftd.get("sample_data", token)
sensor_modality = sd_record["sensor_modality"]
if sensor_modality == "camera":
# Load boxes and image.
data_path, boxes, camera_intrinsic = lyftd.get_sample_data(token, box_vis_level=BoxVisibility.ANY)
data = Image.open(data_path)
# Init axes.
fig, ax = plt.subplots(1, 1, figsize=(9, 16))
# Show image.
ax.imshow(data)
for box in boxes:
c = np.array(LyftDatasetExplorer.get_color(box.name)) / 255.0
box.render(ax, view=camera_intrinsic, normalize=True, colors=(c, c, c))
# Limit visible range.
ax.set_xlim(0, data.size[0])
ax.set_ylim(data.size[1], 0)
fig.savefig("./temp_figs/{}.png".format(token), dpi=300)
plt.close()
def render_sample(sample_token, dataLyft3D):
fig, axes = plt.subplots(1, 1, figsize=(12, 6))
cam = dataLyft3D.get("sample_data", sample_token)
data_path, boxes, camera_intrinsic = dataLyft3D.get_sample_data(
sample_token, selected_anntokens=None)
im = Image.open(data_path)
axes.imshow(im)
axes.set_title(cam["channel"])
axes.axis("off")
axes.set_aspect("equal")
for box in boxes:
c = np.array(LyftDatasetExplorer.get_color(box.name)) / 255.0
render_box(box,
axes,
view=camera_intrinsic,
normalize=True,
colors=(c, c, c),
im=im)
# os.symlink("/media/pzr/软件/kaggle/3D/train_images", 'images')
# os.symlink("/media/pzr/软件/kaggle/3D/train_maps", 'maps')
# os.symlink("/media/pzr/软件/kaggle/3D/train_lidar", 'lidar')
# Our code will generate data, visualization and model checkpoints, they will be persisted to disk in this folder
DATA_PATH = '/media/pzr/软件/kaggle/3D/train_data'
ARTIFACTS_FOLDER = "/media/pzr/软件/kaggle/3D/artifacts/"
# 这里我软链接建立了但是还是链接不上map_raster_palo_alto.png文件，所以改了一下map.json 的最后一行为"filename": "test_maps/map_raster_palo_alto.png", "category": "semantic_prior"}]
# 在LyftDataset源码可以看到里面会读train_data里面的所有json 文件
# 所有修改过的 或者 我添加的代码在前面用pzr表示
level5data = LyftDataset(data_path='.', json_path=DATA_PATH, verbose=True)
os.makedirs(ARTIFACTS_FOLDER, exist_ok=True)
# pzr 这个Explorer 可以输出一些数据信息
dataHelper = LyftDatasetExplorer(level5data)
dataHelper.list_categories()
# Category stats
# animal n= 186, width= 0.36±0.12, len= 0.73±0.19, height= 0.51±0.16, lw_aspect= 2.16±0.56
# bicycle n=20928, width= 0.63±0.24, len= 1.76±0.29, height= 1.44±0.37, lw_aspect= 3.20±1.17
# bus n= 8729, width= 2.96±0.24, len=12.34±3.41, height= 3.44±0.31, lw_aspect= 4.17±1.10
# car n=534911, width= 1.93±0.16, len= 4.76±0.53, height= 1.72±0.24, lw_aspect= 2.47±0.22
# emergency_vehicle n= 132, width= 2.45±0.43, len= 6.52±1.44, height= 2.39±0.59, lw_aspect= 2.66±0.28
# motorcycle n= 818, width= 0.96±0.20, len= 2.35±0.22, height= 1.59±0.16, lw_aspect= 2.53±0.50
# other_vehicle n=33376, width= 2.79±0.30, len= 8.20±1.71, height= 3.23±0.50, lw_aspect= 2.93±0.53
# pedestrian n=24935, width= 0.77±0.14, len= 0.81±0.17, height= 1.78±0.16, lw_aspect= 1.06±0.20
# truck n=14164, width= 2.84±0.32, len=10.24±4.09, height= 3.44±0.62, lw_aspect= 3.56±1.25
dataHelper.list_attributes()
class Lyft(SequenceDataset):
def __init__(self, root, split=None):
self.data = LyftDataset(data_path=root,
json_path=os.path.join(
root,
os.path.basename(os.path.normpath(root))),
verbose=False)
self.explorer = LyftDatasetExplorer(self.data)
super().__init__(root, split)
def _get_sequences(self):
return [(scene["first_sample_token"], "CAM_FRONT")
for scene in self.data.scene]
def _get_samples(self, sequence):
FIRST_TOKEN, CAMERA = sequence
next_token = FIRST_TOKEN
samples = []
while next_token != "":
sample = self.data.get("sample", next_token)
next_token = sample["next"]
image = sample["data"][CAMERA]
samples.append(image)
return samples
def _load_sample(self, seq_i, sample):
sd_record = self.data.get("sample_data", sample)
cs_record = self.data.get("calibrated_sensor",
sd_record["calibrated_sensor_token"])
pose_record = self.data.get("ego_pose", sd_record["ego_pose_token"])
K = np.array(cs_record["camera_intrinsic"]).astype(np.float32)
T = Quaternion(pose_record["rotation"]).transformation_matrix
T[:3, 3] = np.array(pose_record["translation"])
#print("before", T)
T = T @ np.array([
[0, 0, 1, 0],
[-1, 0, 0, 0],
[0, -1, 0, 0],
[0, 0, 0, 1],
])
#print("after", T)
img, K = self._load_image_from_disk(
self.data.get_sample_data_path(sample), K)
return [img, T, K]
def _load_depth(self, sample):
sample_data = self.data.get("sample_data", sample)
LIDAR_TOKEN = self.data.get(
"sample", sample_data["sample_token"])["data"]["LIDAR_TOP"]
points, depths, image = self.explorer.map_pointcloud_to_image(
LIDAR_TOKEN, sample)
W, H = image.size
crop, scale = self.calc_crop(H, W)
dmap = np.zeros(self.imHW, dtype=np.float64)
coords = points[:2, :] * scale
coords[1] -= crop
mask = np.logical_and(coords[1, :] >= 0, coords[1, :] < self.imHW[0])
coords = coords[:, mask].astype(np.int)
depths = depths[mask]
dmap[coords[1], coords[0]] = depths
return dmap