def _try_to_fix_agent_yaw(
sat_rast: SemanticRasterizer,
agent: np.ndarray,
selected_track_id: int,
history_frames: np.ndarray,
history_agents: List[np.ndarray],
history_num_frames: int,
angle_diff_th_in_degree: float,
distance_th_in_agent: float,
num_use_history: int,
lane_polygon_buffer: float,
) -> None:
if PERCEPTION_LABELS[np.argmax(
agent["label_probabilities"])] != "PERCEPTION_LABEL_CAR":
return agent, False
agent_centroid_m = agent["centroid"]
agent_yaw_rad = float(agent["yaw"])
# TODO(kaizaburo): history_positions も world coords で計算する (それに合わせてハイパラも調整する)
world_from_agent = compute_agent_pose(agent_centroid_m, agent_yaw_rad)
agent_from_world = np.linalg.inv(world_from_agent)
# history_num_frames + 1 because it also includes the current frame
history_coords_offset, _, history_availability = _create_targets_for_deep_prediction(
history_num_frames + 1, history_frames, selected_track_id,
history_agents, agent_from_world, agent_yaw_rad)
history_positions = []
for pos, avail in zip(history_coords_offset, history_availability):
if avail == 0.0:
break
history_positions.append(pos)
history_positions = np.array(history_positions)
new_yaw = fix_yaw_by_surrounding_lane_and_history(
sat_rast=sat_rast,
centroid=agent_centroid_m,
yaw_origin=agent_yaw_rad,
history_positions=history_positions,
angle_diff_th_in_degree=angle_diff_th_in_degree,
distance_th_in_agent=distance_th_in_agent,
num_use_history=num_use_history,
lane_polygon_buffer=lane_polygon_buffer,
)
if new_yaw is not None:
agent = agent.copy()
agent["yaw"] = new_yaw
return agent, True
else:
return agent, False
def base_displacement(zarr_dataset: ChunkedDataset, cfg: dict) -> np.ndarray:
future_num_frames = cfg["model_params"]["future_num_frames"]
ref_frame = zarr_dataset.frames[0]
world_from_agent = compute_agent_pose(
ref_frame["ego_translation"][:2],
rotation33_as_yaw((ref_frame["ego_rotation"])))
future_positions, *_ = get_relative_poses(
num_frames=future_num_frames,
frames=zarr_dataset.frames[1:1 + future_num_frames],
selected_track_id=None,
agents=[np.empty(0) for _ in range(future_num_frames)],
agent_from_world=np.linalg.inv(world_from_agent),
current_agent_yaw=rotation33_as_yaw(ref_frame["ego_rotation"]),
)
return future_positions
def generate_agent_sample_tl_persistence(
state_index: int,
frames: np.ndarray,
agents: np.ndarray,
tl_faces: np.ndarray,
selected_track_id: Optional[int],
render_context: RenderContext,
history_num_frames: int,
history_step_size: int,
future_num_frames: int,
future_step_size: int,
filter_agents_threshold: float,
rasterizer: Optional[Rasterizer] = None,
perturbation: Optional[Perturbation] = None,
) -> dict:
"""Generates the inputs and targets to train a deep prediction model. A deep prediction model takes as input
the state of the world (here: an image we will call the "raster"), and outputs where that agent will be some
seconds into the future.
This function has a lot of arguments and is intended for internal use, you should try to use higher level classes
and partials that use this function.
Args:
state_index (int): The anchor frame index, i.e. the "current" timestep in the scene
frames (np.ndarray): The scene frames array, can be numpy array or a zarr array
agents (np.ndarray): The full agents array, can be numpy array or a zarr array
tl_faces (np.ndarray): The full traffic light faces array, can be numpy array or a zarr array
selected_track_id (Optional[int]): Either None for AV, or the ID of an agent that you want to
predict the future of. This agent is centered in the raster and the returned targets are derived from
their future states.
raster_size (Tuple[int, int]): Desired output raster dimensions
pixel_size (np.ndarray): Size of one pixel in the real world
ego_center (np.ndarray): Where in the raster to draw the ego, [0.5,0.5] would be the center
history_num_frames (int): Amount of history frames to draw into the rasters
history_step_size (int): Steps to take between frames, can be used to subsample history frames
future_num_frames (int): Amount of history frames to draw into the rasters
future_step_size (int): Steps to take between targets into the future
filter_agents_threshold (float): Value between 0 and 1 to use as cutoff value for agent filtering
based on their probability of being a relevant agent
rasterizer (Optional[Rasterizer]): Rasterizer of some sort that draws a map image
perturbation (Optional[Perturbation]): Object that perturbs the input and targets, used
to train models that can recover from slight divergence from training set data
Raises:
ValueError: A ValueError is returned if the specified ``selected_track_id`` is not present in the scene
or was filtered by applying the ``filter_agent_threshold`` probability filtering.
Returns:
dict: a dict object with the raster array, the future offset coordinates (meters),
the future yaw angular offset, the future_availability as a binary mask
"""
# the history slice is ordered starting from the latest frame and goes backward in time., ex. slice(100, 91, -2)
all_history_slice = get_history_slice(state_index, state_index, history_step_size, include_current_state=True)
history_slice = get_history_slice(state_index, history_num_frames, history_step_size, include_current_state=True)
future_slice = get_future_slice(state_index, future_num_frames, future_step_size)
all_history_frames = frames[all_history_slice].copy() # TL data will be based on all history
history_frames = frames[history_slice].copy() # copy() required if the object is a np.ndarray
future_frames = frames[future_slice].copy()
sorted_frames = np.concatenate((history_frames[::-1], future_frames)) # from past to future
# get agents (past and future)
agent_slice = get_agents_slice_from_frames(sorted_frames[0], sorted_frames[-1])
agents = agents[agent_slice].copy() # this is the minimum slice of agents we need
history_frames["agent_index_interval"] -= agent_slice.start # sync interval with the agents array
future_frames["agent_index_interval"] -= agent_slice.start # sync interval with the agents array
history_agents = filter_agents_by_frames(history_frames, agents)
future_agents = filter_agents_by_frames(future_frames, agents)
# sync interval with the traffic light faces array
tl_slice = get_tl_faces_slice_from_frames(all_history_frames[-1], all_history_frames[0]) # -1 is the farthest
all_history_frames["traffic_light_faces_index_interval"] -= tl_slice.start
history_tl_faces = filter_tl_faces_by_frames(all_history_frames, tl_faces[tl_slice].copy())
# State you want to predict the future of.
cur_frame = history_frames[0]
cur_agents = history_agents[0]
if selected_track_id is None:
agent_centroid_m = cur_frame["ego_translation"][:2]
agent_yaw_rad = rotation33_as_yaw(cur_frame["ego_rotation"])
agent_extent_m = np.asarray((EGO_EXTENT_LENGTH, EGO_EXTENT_WIDTH, EGO_EXTENT_HEIGHT))
selected_agent = None
else:
# this will raise IndexError if the agent is not in the frame or under agent-threshold
# this is a strict error, we cannot recover from this situation
try:
agent = filter_agents_by_track_id(
filter_agents_by_labels(cur_agents, filter_agents_threshold), selected_track_id
)[0]
except IndexError:
raise ValueError(f" track_id {selected_track_id} not in frame or below threshold")
agent_centroid_m = agent["centroid"]
agent_yaw_rad = float(agent["yaw"])
agent_extent_m = agent["extent"]
selected_agent = agent
input_im = (
None
if not rasterizer
else rasterizer.rasterize(history_frames, history_agents, history_tl_faces, selected_agent)
)
world_from_agent = compute_agent_pose(agent_centroid_m, agent_yaw_rad)
agent_from_world = np.linalg.inv(world_from_agent)
raster_from_world = render_context.raster_from_world(agent_centroid_m, agent_yaw_rad)
future_coords_offset, future_yaws_offset, future_availability = _create_targets_for_deep_prediction(
future_num_frames, future_frames, selected_track_id, future_agents, agent_from_world, agent_yaw_rad
)
# history_num_frames + 1 because it also includes the current frame
history_coords_offset, history_yaws_offset, history_availability = _create_targets_for_deep_prediction(
history_num_frames + 1, history_frames, selected_track_id, history_agents, agent_from_world, agent_yaw_rad
)
return {
"image": input_im,
"target_positions": future_coords_offset,
"target_yaws": future_yaws_offset,
"target_availabilities": future_availability,
"history_positions": history_coords_offset,
"history_yaws": history_yaws_offset,
"history_availabilities": history_availability,
"world_to_image": raster_from_world, # TODO deprecate
"raster_from_agent": raster_from_world @ world_from_agent,
"raster_from_world": raster_from_world,
"agent_from_world": agent_from_world,
"world_from_agent": world_from_agent,
"centroid": agent_centroid_m,
"yaw": agent_yaw_rad,
"extent": agent_extent_m,
}
def generate_multi_agent_sample(
state_index: int,
frames: np.ndarray,
agents: np.ndarray,
tl_faces: np.ndarray,
selected_track_id: Optional[int],
render_context: RenderContext,
history_num_frames: int,
history_step_size: int,
future_num_frames: int,
future_step_size: int,
filter_agents_threshold: float,
rasterizer: Optional[Rasterizer] = None,
perturbation: Optional[Perturbation] = None,
min_frame_history: int = MIN_FRAME_HISTORY,
min_frame_future: int = MIN_FRAME_FUTURE,
) -> dict:
"""Generates the inputs and targets to train a deep prediction model. A deep prediction model takes as input
the state of the world (here: an image we will call the "raster"), and outputs where that agent will be some
seconds into the future.
This function has a lot of arguments and is intended for internal use, you should try to use higher level classes
and partials that use this function.
Args:
state_index (int): The anchor frame index, i.e. the "current" timestep in the scene
frames (np.ndarray): The scene frames array, can be numpy array or a zarr array
agents (np.ndarray): The full agents array, can be numpy array or a zarr array
tl_faces (np.ndarray): The full traffic light faces array, can be numpy array or a zarr array
selected_track_id (Optional[int]): Either None for AV, or the ID of an agent that you want to
predict the future of. This agent is centered in the raster and the returned targets are derived from
their future states.
render_context (RenderContext):
raster_size (Tuple[int, int]): Desired output raster dimensions
pixel_size (np.ndarray): Size of one pixel in the real world
ego_center (np.ndarray): Where in the raster to draw the ego, [0.5,0.5] would be the center
history_num_frames (int): Amount of history frames to draw into the rasters
history_step_size (int): Steps to take between frames, can be used to subsample history frames
future_num_frames (int): Amount of history frames to draw into the rasters
future_step_size (int): Steps to take between targets into the future
filter_agents_threshold (float): Value between 0 and 1 to use as cutoff value for agent filtering
based on their probability of being a relevant agent
rasterizer (Optional[Rasterizer]): Rasterizer of some sort that draws a map image
perturbation (Optional[Perturbation]): Object that perturbs the input and targets, used
to train models that can recover from slight divergence from training set data
Raises:
ValueError: A ValueError is returned if the specified ``selected_track_id`` is not present in the scene
or was filtered by applying the ``filter_agent_threshold`` probability filtering.
Returns:
dict: a dict object with the raster array, the future offset coordinates (meters),
the future yaw angular offset, the future_availability as a binary mask
"""
# the history slice is ordered starting from the latest frame and goes backward in time., ex. slice(100, 91, -2)
history_slice = get_history_slice(state_index,
history_num_frames,
history_step_size,
include_current_state=True)
future_slice = get_future_slice(state_index, future_num_frames,
future_step_size)
history_frames = frames[history_slice].copy(
) # copy() required if the object is a np.ndarray
future_frames = frames[future_slice].copy()
sorted_frames = np.concatenate(
(history_frames[::-1], future_frames)) # from past to future
# get agents (past and future)
agent_slice = get_agents_slice_from_frames(sorted_frames[0],
sorted_frames[-1])
agents = agents[agent_slice].copy(
) # this is the minimum slice of agents we need
history_frames[
"agent_index_interval"] -= agent_slice.start # sync interval with the agents array
future_frames[
"agent_index_interval"] -= agent_slice.start # sync interval with the agents array
history_agents = filter_agents_by_frames(history_frames, agents)
future_agents = filter_agents_by_frames(future_frames, agents)
try:
tl_slice = get_tl_faces_slice_from_frames(
history_frames[-1], history_frames[0]) # -1 is the farthest
# sync interval with the traffic light faces array
history_frames["traffic_light_faces_index_interval"] -= tl_slice.start
history_tl_faces = filter_tl_faces_by_frames(history_frames,
tl_faces[tl_slice].copy())
except ValueError:
history_tl_faces = [
np.empty(0, dtype=TL_FACE_DTYPE) for _ in history_frames
]
if perturbation is not None:
history_frames, future_frames = perturbation.perturb(
history_frames=history_frames, future_frames=future_frames)
# State you want to predict the future of.
cur_frame = history_frames[0]
cur_agents = history_agents[0]
cur_agents = filter_agents_by_labels(cur_agents, filter_agents_threshold)
agent_track_ids_u64 = cur_agents["track_id"]
# uint64 --> int64
agent_track_ids = agent_track_ids_u64.astype(np.int64)
assert np.alltrue(agent_track_ids == agent_track_ids_u64)
agent_track_ids = np.concatenate(
[np.array([-1], dtype=np.int64), agent_track_ids])
# Draw image with Ego car in center
selected_agent = None
input_im = (None if not rasterizer else rasterizer.rasterize(
history_frames, history_agents, history_tl_faces, selected_agent))
future_coords_offset_list = []
future_yaws_offset_list = []
future_availability_list = []
history_coords_offset_list = []
history_yaws_offset_list = []
history_availability_list = []
agent_centroid_list = []
agent_yaw_list = []
agent_extent_list = []
filtered_track_ids_list = []
for selected_track_id in agent_track_ids:
if selected_track_id == -1:
agent_centroid = cur_frame["ego_translation"][:2]
agent_yaw_rad = rotation33_as_yaw(cur_frame["ego_rotation"])
agent_extent = np.asarray(
(EGO_EXTENT_LENGTH, EGO_EXTENT_WIDTH, EGO_EXTENT_HEIGHT))
world_from_agent = compute_agent_pose(agent_centroid,
agent_yaw_rad)
agent_from_world = np.linalg.inv(world_from_agent)
raster_from_world = render_context.raster_from_world(
agent_centroid, agent_yaw_rad)
agent_origin = np.zeros((2, ), dtype=np.float32)
else:
# this will raise IndexError if the agent is not in the frame or under agent-threshold
# this is a strict error, we cannot recover from this situation
try:
agent = filter_agents_by_track_id(cur_agents,
selected_track_id)[0]
except IndexError:
raise ValueError(
f" track_id {selected_track_id} not in frame or below threshold"
)
agent_centroid = agent["centroid"]
agent_yaw_rad = agent["yaw"]
agent_extent = agent["extent"]
agent_origin = transform_point(agent_centroid, agent_from_world)
future_coords_offset, future_yaws_offset, future_availability = _create_targets_for_deep_prediction(
future_num_frames, future_frames, selected_track_id, future_agents,
agent_from_world, agent_yaw_rad, agent_origin)
if selected_track_id != -1 and np.sum(
future_availability) < min_frame_future:
# Not enough future to predict, skip this agent.
continue
# history_num_frames + 1 because it also includes the current frame
history_coords_offset, history_yaws_offset, history_availability = _create_targets_for_deep_prediction(
history_num_frames + 1, history_frames, selected_track_id,
history_agents, agent_from_world, agent_yaw_rad, agent_origin)
if selected_track_id != -1 and np.sum(
history_availability) < min_frame_history:
# Not enough history to predict, skip this agent.
continue
future_coords_offset_list.append(future_coords_offset)
future_yaws_offset_list.append(future_yaws_offset)
future_availability_list.append(future_availability)
history_coords_offset_list.append(history_coords_offset)
history_yaws_offset_list.append(history_yaws_offset)
history_availability_list.append(history_availability)
agent_centroid_list.append(agent_centroid)
agent_yaw_list.append(agent_yaw_rad)
agent_extent_list.append(agent_extent)
filtered_track_ids_list.append(selected_track_id)
# Get pixel coordinate
agent_centroid_array = np.array(agent_centroid_list)
agent_centroid_in_pixel = transform_points(agent_centroid_array,
raster_from_world)
return {
"image": input_im, # (h, w, ch)
# --- All below is in world coordinate ---
"target_positions":
np.array(future_coords_offset_list), # (n_agents, num_frames, 2)
"target_yaws":
np.array(future_yaws_offset_list), # (n_agents, num_frames, 1)
"target_availabilities":
np.array(future_availability_list), # (n_agents, num_frames)
"history_positions":
np.array(history_coords_offset_list), # (n_agents, num_frames, 2)
"history_yaws":
np.array(history_yaws_offset_list), # (n_agents, num_frames, 1)
"history_availabilities":
np.array(history_availability_list), # (n_agents, num_frames)
# "world_to_image": raster_from_world, # (3, 3)
"raster_from_world": raster_from_world, # (3, 3)
"centroid": agent_centroid_array, # (n_agents, 2)
"yaw": np.array(agent_yaw_list), # (n_agents, 1)
"extent": np.array(agent_extent_list), # (n_agents, 3)
"track_ids": np.array(filtered_track_ids_list), # (n_agents)
"centroid_pixel": agent_centroid_in_pixel, # (n_agents, 2)
}
