def generate_prediction_dict(input_dict, index, record_dict):
# finally generate predictions.
sample_idx = input_dict['sample_idx'][
index] if 'sample_idx' in input_dict else -1
boxes3d_lidar_preds = record_dict['boxes'].cpu().numpy()
if boxes3d_lidar_preds.shape[0] == 0:
return {'sample_idx': sample_idx}
calib = input_dict['calib'][index]
image_shape = input_dict['image_shape'][index]
boxes3d_camera_preds = box_utils.boxes3d_lidar_to_camera(
boxes3d_lidar_preds, calib)
boxes2d_image_preds = box_utils.boxes3d_camera_to_imageboxes(
boxes3d_camera_preds, calib, image_shape=image_shape)
# predictions
predictions_dict = {
'bbox': boxes2d_image_preds,
'box3d_camera': boxes3d_camera_preds,
'box3d_lidar': boxes3d_lidar_preds,
'scores': record_dict['scores'].cpu().numpy(),
'label_preds': record_dict['labels'].cpu().numpy(),
'sample_idx': sample_idx,
}
return predictions_dict
def generate_prediction_dict(input_dict, index, record_dict):
"""
Generate the prediction dict for EACH sample, called by the post processing. All this fn
does really is mapping prediciton boxes to camera coordinates putting things together in
a predictions_dict.
Args:
input_dict: provided by the dataset to provide dataset-specific information
index: batch index of current sample
record_dict: the predicted results of current sample from the detector,
which currently includes these keys: {
'boxes': (N, 7 + C) [x, y, z, w, l, h, heading_in_kitti] in LiDAR coords
'scores': (N)
'labels': (N）
}
Returns:
predictions_dict: the required prediction dict of current scene for specific dataset
"""
# finally generate predictions.
sample_idx = input_dict['sample_idx'][
index] if 'sample_idx' in input_dict else -1
boxes3d_lidar_preds = record_dict['boxes'].cpu().numpy()
if boxes3d_lidar_preds.shape[0] == 0:
return {'sample_idx': sample_idx}
calib = input_dict['calib'][index]
image_shape = input_dict['image_shape'][index]
boxes3d_camera_preds = box_utils.boxes3d_lidar_to_camera(
boxes3d_lidar_preds, calib)
boxes2d_image_preds = box_utils.boxes3d_camera_to_imageboxes(
boxes3d_camera_preds, calib, image_shape=image_shape)
# predictions
predictions_dict = {
'bbox': boxes2d_image_preds,
'box3d_camera': boxes3d_camera_preds,
'box3d_lidar': boxes3d_lidar_preds,
'scores': record_dict['scores'].cpu().numpy(),
'label_preds': record_dict['labels'].cpu().numpy(),
'sample_idx': sample_idx,
}
return predictions_dict
def generate_prediction_dict(input_dict, index, record_dict):
# finally generate predictions.
sample_idx = input_dict['sample_idx'][
index] if 'sample_idx' in input_dict else -1
try:
boxes3d_lidar_preds = record_dict['boxes'].cpu().numpy()
except:
boxes3d_lidar_preds = record_dict['boxes'].detach().cpu().numpy()
if boxes3d_lidar_preds.shape[0] == 0:
return {'sample_idx': sample_idx}
if not cfg.DATA_CONFIG.TS_DATA:
calib = input_dict['calib'][index]
image_shape = input_dict['image_shape'][index]
boxes3d_camera_preds = box_utils.boxes3d_lidar_to_camera(
boxes3d_lidar_preds, calib)
boxes2d_image_preds = box_utils.boxes3d_camera_to_imageboxes(
boxes3d_camera_preds, calib, image_shape=image_shape)
# predictions
predictions_dict = {
'bbox': boxes2d_image_preds,
'box3d_camera': boxes3d_camera_preds,
'box3d_lidar': boxes3d_lidar_preds,
'scores': record_dict['scores'].cpu().numpy(),
'label_preds': record_dict['labels'].cpu().numpy(),
'sample_idx': sample_idx,
}
else:
boxes3d_camera_preds = box_utils.ts_boxes3d_lidar_to_camera(
boxes3d_lidar_preds)
predictions_dict = {
'bbox': np.zeros((boxes3d_lidar_preds.shape[0], 4)),
'box3d_camera': boxes3d_camera_preds,
'box3d_lidar': boxes3d_lidar_preds,
'scores': record_dict['scores'].detach().cpu().numpy(),
'label_preds': record_dict['labels'].detach().cpu().numpy(),
'sample_idx': sample_idx,
}
return predictions_dict