def _load_transformed_wall_maps(self, scene_map_info, episode):
seen_maps = []
wall_maps = []
start_position = episode.start_position # (X, Y, Z)
start_rotation = quaternion_xyzw_to_wxyz(episode.start_rotation)
start_heading = compute_heading_from_quaternion(start_rotation)
for floor_data in scene_map_info:
seen_map = np.load(floor_data["seen_map_path"])
wall_map = np.load(floor_data["wall_map_path"])
# ===== Transform the maps relative to the episode start pose =====
map_view_position = floor_data["world_position"]
map_view_heading = floor_data["world_heading"]
# Originally, Z is downward and X is rightward.
# Convert it to X upward and Y rightward
x_map, y_map = -map_view_position[2], map_view_position[0]
theta_map = map_view_heading
x_start, y_start = -start_position[2], start_position[0]
theta_start = start_heading
# Compute relative coordinates
r_rel = math.sqrt((x_start - x_map) ** 2 + (y_start - y_map) ** 2)
phi_rel = math.atan2(y_start - y_map, x_start - x_map) - theta_map
x_rel = r_rel * math.cos(phi_rel) / self.config.MAP_SCALE
y_rel = r_rel * math.sin(phi_rel) / self.config.MAP_SCALE
theta_rel = theta_start - theta_map
# Convert these to image coordinates with X being rightward and Y
# being downward
x_img_rel = y_rel
y_img_rel = -x_rel
theta_img_rel = theta_rel
x_trans = torch.Tensor([[x_img_rel, y_img_rel, theta_img_rel]])
# Perform the transformations
p_seen_map = rearrange(torch.Tensor(seen_map), "h w c -> () c h w")
p_wall_map = rearrange(torch.Tensor(wall_map), "h w c -> () c h w")
p_seen_map_trans = spatial_transform_map(p_seen_map, x_trans)
p_wall_map_trans = spatial_transform_map(p_wall_map, x_trans)
seen_map_trans = asnumpy(p_seen_map_trans)
seen_map_trans = rearrange(seen_map_trans, "() c h w -> h w c")
wall_map_trans = asnumpy(p_wall_map_trans)
wall_map_trans = rearrange(wall_map_trans, "() c h w -> h w c")
seen_maps.append(seen_map_trans)
wall_maps.append(wall_map_trans)
return seen_maps, wall_maps
def _get_mesh_occupancy(self):
agent_position = self.current_episode.start_position
agent_rotation = quaternion_xyzw_to_wxyz(self.current_episode.start_rotation)
a_x, a_y = maps.to_grid(
agent_position[0],
agent_position[2],
self._coordinate_min,
self._coordinate_max,
self._map_resolution,
)
# The map size here represents size around the agent, not infront.
mrange = int(self.map_size * 1.5 / 2.0)
# Add extra padding if map range is coordinates go out of bounds
y_start = a_y - mrange
y_end = a_y + mrange
x_start = a_x - mrange
x_end = a_x + mrange
x_l_pad, y_l_pad, x_r_pad, y_r_pad = 0, 0, 0, 0
H, W = self._top_down_map.shape
if x_start < 0:
x_l_pad = int(-x_start)
x_start += x_l_pad
x_end += x_l_pad
if x_end >= W:
x_r_pad = int(x_end - W + 1)
if y_start < 0:
y_l_pad = int(-y_start)
y_start += y_l_pad
y_end += y_l_pad
if y_end >= H:
y_r_pad = int(y_end - H + 1)
ego_map = np.pad(self._top_down_map, ((y_l_pad, y_r_pad), (x_l_pad, x_r_pad)))
ego_map = ego_map[y_start : (y_end + 1), x_start : (x_end + 1)]
if ego_map.shape[0] == 0 or ego_map.shape[1] == 0:
ego_map = np.zeros((2 * mrange + 1, 2 * mrange + 1), dtype=np.uint8)
# Rotate to get egocentric map
# Negative since the value returned is clockwise rotation about Y,
# but we need anti-clockwise rotation
agent_heading = -compute_heading_from_quaternion(agent_rotation)
agent_heading = math.degrees(agent_heading)
half_size = ego_map.shape[0] // 2
center = (half_size, half_size)
M = cv2.getRotationMatrix2D(center, agent_heading, scale=1.0)
ego_map = cv2.warpAffine(
ego_map,
M,
(ego_map.shape[1], ego_map.shape[0]),
flags=cv2.INTER_NEAREST,
borderMode=cv2.BORDER_CONSTANT,
borderValue=(1,),
)
ego_map = ego_map.astype(np.float32)
mrange = int(self.map_size / 2.0)
start_coor = half_size - mrange
end_coor = int(start_coor + self.map_size - 1)
ego_map = ego_map[start_coor : (end_coor + 1), start_coor : (end_coor + 1)]
# This map is currently 0 if occupied and 1 if unoccupied. Flip it.
ego_map = 1.0 - ego_map
# Flip the x axis because to_grid() flips the conventions
ego_map = np.flip(ego_map, axis=1)
# Append explored status in the 2nd channel
ego_map = np.stack([ego_map, np.ones_like(ego_map)], axis=2)
return ego_map
def _get_wall_occupancy(self, episode, agent_state):
episode_id = (episode.episode_id, episode.scene_id)
# Load the episode specific maps only if the episode has changed
if self.current_wall_episode_id != episode_id:
self.current_wall_episode_id = episode_id
if self.config.GT_TYPE == "wall_occupancy":
scene_id = episode.scene_id.split("/")[-1]
self._scene_maps_info = self._all_maps_info[scene_id]
# Load the maps per floor
seen_maps, wall_maps = self._load_transformed_wall_maps(
self._scene_maps_info, episode,
)
self._scene_maps = {}
self._scene_maps["seen_maps"] = seen_maps
self._scene_maps["wall_maps"] = wall_maps
agent_state = self._sim.get_agent_state()
current_height = agent_state.position[1]
best_floor_idx = None
best_floor_dist = math.inf
for floor_idx, floor_data in enumerate(self._scene_maps_info):
floor_height = floor_data["floor_height"]
if abs(current_height - floor_height) < best_floor_dist:
best_floor_idx = floor_idx
best_floor_dist = abs(current_height - floor_height)
assert best_floor_idx is not None
current_wall_map = self._scene_maps["wall_maps"][best_floor_idx]
# Take only channel 0 as both channels have save values
current_wall_map = current_wall_map[..., 0]
# ========= Get local egocentric crop of the current wall map =========
# Compute relative pose of agent from start location
start_position = episode.start_position # (X, Y, Z)
start_rotation = quaternion_xyzw_to_wxyz(episode.start_rotation)
start_heading = compute_heading_from_quaternion(start_rotation)
start_pose = torch.Tensor(
[[-start_position[2], start_position[0], start_heading]]
)
agent_position = agent_state.position
agent_heading = compute_heading_from_quaternion(agent_state.rotation)
agent_pose = torch.Tensor(
[[-agent_position[2], agent_position[0], agent_heading]]
)
rel_pose = subtract_pose(start_pose, agent_pose)[0] # (3,)
# Compute agent position on the map image
map_scale = self.config.MAP_SCALE
H, W = current_wall_map.shape[:2]
Hby2, Wby2 = (H + 1) // 2, (W + 1) // 2
agent_map_x = int(rel_pose[1].item() / map_scale + Wby2)
agent_map_y = int(-rel_pose[0].item() / map_scale + Hby2)
# Crop the region around the agent.
mrange = int(1.5 * self.map_size)
# Add extra padding if map range is coordinates go out of bounds
y_start = agent_map_y - mrange
y_end = agent_map_y + mrange
x_start = agent_map_x - mrange
x_end = agent_map_x + mrange
x_l_pad, y_l_pad, x_r_pad, y_r_pad = 0, 0, 0, 0
H, W = current_wall_map.shape
if x_start < 0:
x_l_pad = int(-x_start)
x_start += x_l_pad
x_end += x_l_pad
if x_end >= W:
x_r_pad = int(x_end - W + 1)
if y_start < 0:
y_l_pad = int(-y_start)
y_start += y_l_pad
y_end += y_l_pad
if y_end >= H:
y_r_pad = int(y_end - H + 1)
ego_map = np.pad(current_wall_map, ((y_l_pad, y_r_pad), (x_l_pad, x_r_pad)))
ego_map = ego_map[y_start: (y_end + 1), x_start: (x_end + 1)]
agent_heading = rel_pose[2].item()
agent_heading = math.degrees(agent_heading)
half_size = ego_map.shape[0] // 2
center = (half_size, half_size)
M = cv2.getRotationMatrix2D(center, agent_heading, scale=1.0)
ego_map = cv2.warpAffine(
ego_map,
M,
(ego_map.shape[1], ego_map.shape[0]),
flags=cv2.INTER_NEAREST,
borderMode=cv2.BORDER_CONSTANT,
borderValue=(1,),
)
ego_map = ego_map.astype(np.float32)
mrange = int(self.map_size)
ego_map = ego_map[
(half_size - mrange): (half_size + mrange),
(half_size - mrange): (half_size + mrange),
]
# Get forward region infront of the agent
half_size = ego_map.shape[0] // 2
quarter_size = ego_map.shape[0] // 4
center = (half_size, half_size)
ego_map = ego_map[0:half_size, quarter_size: (quarter_size + half_size)]
# Append explored status in the 2nd channel
ego_map = np.stack([ego_map, ego_map], axis=2)
return ego_map