Currently we only provide the processing of Waymo dataset.
-
please place the soft link of the data folder based on the following structure:
detection ├── data │ ├── waymo │ │ │── ImageSets │ │ │── raw_data │ │ │ │── segment-xxxxxxxx.tfrecord │ │ │ │── .... ├── detzero_det ├── tools -
process waymo infos:
cd detection python -m detzero_det.datasets.waymo.waymo_preprocess --cfg_file tools/cfgs/det_dataset_cfgs/waymo_one_sweep.yaml --func create_waymo_infos -
generate database for gt-sampling
cd detection python -m detzero_det.datasets.waymo.waymo_preprocess --cfg_file tools/cfgs/det_dataset_cfgs/waymo_one_sweep.yaml --func create_waymo_database
We have provided a flexible data info structure and processing logic to satisfy single-frame or multi-frame (e.g., 2, 3, 5, ...) point clouds loading without further repeated pre-processings.
def get_sweep_idxs(current_info, sweep_count=[0, 0], current_idx=0):
assert type(sweep_count) is list and len(sweep_count) == 2,\
"Please give the upper and lower range of frames you want to process!"
current_sample_idx = current_info["sample_idx"]
current_seq_len = current_info["sequence_len"]
target_sweep_list = np.array(list(range(sweep_count[0], sweep_count[1]+1)))
target_sample_list = current_sample_idx + target_sweep_list
# set the low and high thresh to extract multi frames in current sequence
target_sample_list = [i if i >= 0 else 0 for i in target_sample_list]
target_sample_list = [i if i < current_seq_len else current_seq_len-1 for i in target_sample_list]
# get the index of target frames in the waymo info list
target_idx_res = np.array(target_sample_list) - current_sample_idx
target_idx_list = current_idx + target_idx_res
return target_idx_list