Before downloading or using the NuScenes dataset, please follow these important steps:
- Visit the official NuScenes website.
- Register for an account if you haven't already done so.
- Carefully read and agree to the NuScenes terms of use.
Ensure you have completed these steps before proceeding with the dataset download and use.
Download the raw data from the official NuScenes website. Then, create directories for NuScenes data and optionally create a symbolic link if you have the data elsewhere.
mkdir -p ./data/nuscenes
ln -s $PATH_TO_NUSCENES ./data/nuscenes/raw # ['v1.0-mini', 'v1.0-trainval', 'v1.0-test'] lies in it
We'll use the v1.0-mini split in our examples. The process is similar for other splits.
pip install nuscenes-devkit
To process the 10 scenes in NuScenes v1.0-mini split, you can run:
# export PYTHONPATH=\path\to\project
python datasets/preprocess.py \
--data_root data/nuscenes/raw \
--target_dir data/nuscenes/processed \
--dataset nuscenes \
--split v1.0-mini \
--start_idx 0 \
--num_scenes 10 \
--interpolate_N 4 \
--workers 32 \
--process_keys images lidar calib dynamic_masks objects
The extracted data will be stored in the data/nuscenes/processed_10Hz
directory.
interpolate_N
: Increases frame rate by interpolating between keyframes.
- NuScenes provides synchronized keyframes at
2Hz
. Our script allows interpolation to increase up to10Hz
. interpolate_N = 4
: Interpolates 4 frames between original synchronized keyframes.- Result:
10Hz
frame rate((4 + 1) * 2 Hz)
- Note: We recommend using
interpolate_N = 4
. Whileinterpolate_N = 5 (12 Hz)
is possible, it may lead to frame drop issues. Although the camera captures at12 Hz
, occasional frame misses during recording can cause data gaps at higher interpolation rates.
To generate:
- sky masks (required)
- fine dynamic masks (optional)
Follow these steps:
mmcv-full=1.2.7
, which relies on pytorch=1.8
(pytorch<1.9). Hence, a separate conda env is required.
#-- Set conda env
conda create -n segformer python=3.8
conda activate segformer
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
#-- Install mmcv-full
pip install timm==0.3.2 pylint debugpy opencv-python-headless attrs ipython tqdm imageio scikit-image omegaconf
pip install mmcv-full==1.2.7 --no-cache-dir
#-- Clone and install segformer
git clone https://github.com/NVlabs/SegFormer
cd SegFormer
pip install .
Download the pretrained model segformer.b5.1024x1024.city.160k.pth
from the google_drive / one_drive links in https://github.com/NVlabs/SegFormer#evaluation .
Remember the location where you download into, and pass it to the script in the next step with --checkpoint
.
conda activate segformer
segformer_path=/path/to/segformer
split=mini
python datasets/tools/extract_masks.py \
--data_root data/nuscenes/processed_10Hz/$split \
--segformer_path=$segformer_path \
--checkpoint=$segformer_path/pretrained/segformer.b5.1024x1024.city.160k.pth \
--start_idx 0 \
--num_scenes 10 \
--process_dynamic_mask
Replace /path/to/segformer
with the actual path to your Segformer installation.
Note: The --process_dynamic_mask
flag is included to process fine dynamic masks along with sky masks.
This process will extract the required masks from your processed data.
To utilize the SMPL-Gaussian to model pedestrians, please first download the SMPL models.
- Download SMPL v1.1 (
SMPL_python_v.1.1.0.zip
) from the SMPL official website - Move
SMPL_python_v.1.1.0/smpl/models/basicmodel_neutral_lbs_10_207_0_v1.1.0.pkl
toPROJECT_ROOT/smpl_models/SMPL_NEUTRAL.pkl
SMPL-Nodes (SMPL-Gaussian Representation) requires Human Body Pose Sequences of pedestrians. We've developed a human body pose processing pipeline for in-the-wild driving data to generate this information. There are two ways to obtain these data:
We have uploaded preprocessed human pose data for v1.0-mini split of NuScenes scenes to Google Drive. You can download and unzip these files without installing any additional environment.
# https://drive.google.com/file/d/1Z0gJVRtPnjvusQVaW7ghZnwfycZStCZx/view?usp=drive_link
# filename: nuscenes_preprocess_humanpose.zip
cd data
gdown 1Z0gJVRtPnjvusQVaW7ghZnwfycZStCZx
unzip nuscenes_preprocess_humanpose.zip
rm nuscenes_preprocess_humanpose.zip
To process human body poses for other NuScenes scenes or to run the processing pipeline yourself, follow the instructions in our Human Pose Processing Guide.
After completing all preprocessing steps, the project files should be organized according to the following structure:
ProjectPath/data/
└── nuscenes/
├── raw/
│ └── [original NuScenes structure]
└── processed_10Hz/
└── mini/
├── 001/
│ ├──images/ # Images: {timestep:03d}_{cam_id}.jpg
│ ├──lidar/ # LiDAR data: {timestep:03d}.bin
│ ├──lidar_pose/ # Lidar poses: {timestep:03d}.txt
│ ├──extrinsics/ # Camera extrinsics: {cam_id}.txt
│ ├──intrinsics/ # Camera intrinsics: {cam_id}.txt
│ ├──sky_masks/ # Sky masks: {timestep:03d}_{cam_id}.png
│ ├──dynamic_masks/ # Dynamic masks: {timestep:03d}_{cam_id}.png
│ ├──fine_dynamic_masks/ # (Optional) Fine dynamic masks: {timestep:03d}_{cam_id}.png
│ ├──objects/ # Object information
│ └──humanpose/ # Preprocessed human body pose: smpl.pkl
├── 002/
└── ...