def get_drivable_area_compliance(
forecasted_trajectories: Dict[int, List[np.ndarray]],
city_names: Dict[int, str],
max_n_guesses: int,
) -> float:
"""Compute drivable area compliance metric.
Args:
forecasted_trajectories: Predicted top-k trajectory dict with key as seq_id and value as list of trajectories.
Each element of the list is of shape (pred_len x 2).
city_names: Dict mapping sequence id to city name.
max_n_guesses: Maximum number of guesses allowed.
Returns:
Mean drivable area compliance
"""
avm = ArgoverseMap()
dac_score = []
for seq_id, trajectories in forecasted_trajectories.items():
city_name = city_names[seq_id]
num_dac_trajectories = 0
n_guesses = min(max_n_guesses, len(trajectories))
for trajectory in trajectories[:n_guesses]:
raster_layer = avm.get_raster_layer_points_boolean(trajectory, city_name, "driveable_area")
if np.sum(raster_layer) == raster_layer.shape[0]:
num_dac_trajectories += 1
dac_score.append(num_dac_trajectories / n_guesses)
return sum(dac_score) / len(dac_score)
def filter_objs_to_roi(instances: np.ndarray, avm: ArgoverseMap,
city_SE3_egovehicle: SE3, city_name: str) -> np.ndarray:
"""Filter objects to the region of interest (5 meter dilation of driveable area).
We ignore instances outside of region of interest (ROI) during evaluation.
Args:
instances: Numpy array of shape (N,) with ObjectLabelRecord entries
avm: Argoverse map object
city_SE3_egovehicle: pose of egovehicle within city map at time of sweep
city_name: name of city where log was captured
Returns:
instances_roi: objects with any of 4 cuboid corners located within ROI
"""
# for each cuboid, get its 4 corners in the egovehicle frame
corners_egoframe = np.vstack([dt.as_2d_bbox() for dt in instances])
corners_cityframe = city_SE3_egovehicle.transform_point_cloud(
corners_egoframe)
corner_within_roi = avm.get_raster_layer_points_boolean(
corners_cityframe, city_name, "roi")
# check for each cuboid if any of its 4 corners lies within the ROI
is_within_roi = corner_within_roi.reshape(-1, 4).any(axis=1)
instances_roi = instances[is_within_roi]
return instances_roi
def get_pruned_guesses(
forecasted_trajectories: Dict[int, List[np.ndarray]],
city_names: Dict[int, str],
gt_trajectories: Dict[int, np.ndarray],
) -> Dict[int, List[np.ndarray]]:
"""Prune the number of guesses using map.
Args:
forecasted_trajectories: Trajectories forecasted by the algorithm.
city_names: Dict mapping sequence id to city name.
gt_trajectories: Ground Truth trajectories.
Returns:
Pruned number of forecasted trajectories.
"""
args = parse_arguments()
avm = ArgoverseMap()
pruned_guesses = {}
for seq_id, trajectories in forecasted_trajectories.items():
city_name = city_names[seq_id]
da_points = []
for trajectory in trajectories:
raster_layer = avm.get_raster_layer_points_boolean(
trajectory, city_name, "driveable_area")
da_points.append(np.sum(raster_layer))
sorted_idx = np.argsort(da_points)[::-1]
pruned_guesses[seq_id] = [
trajectories[i] for i in sorted_idx[:args.prune_n_guesses]
]
return pruned_guesses