Skip to content

Latest commit

 

History

History
271 lines (185 loc) · 10.3 KB

README.md

File metadata and controls

271 lines (185 loc) · 10.3 KB

LidarDM: Generative LiDAR Simulation in a Generated World

Waymax Generation KITTI Generation KITTI Generation KITTI Generation

alt text

Setup

1. Setup Conda Environments

The project was tested with Python 3.10.13 and all the package requirements are listed in REQUIREMENTS.md. To make things easier, we have provided a script to install the enviroment:

sh install_lidardm.sh

2. Datasets Download/Preprocess/GT SDF Generation

LidarDM supports 2 datasets:

  • KITTI-360 for unconditional generation.
  • Waymo Open Dataset for conditional multi-frame generation and downstream tasks.

For each of the datasets, we provide a preprocess to filter out dynamic objects for every scan.

KITTI-360

  1. Download the KITTI-360 Dataset into ./_datasets/kitti/. Make sure to download all of Raw Velodyne Scans (data_3d_raw), 3D Bounding Boxes (data_3d_bboxes), Vehicle Poses (data_poses)

  2. Run the preprocess script:

    python scripts/dataset/preprocess/kitti360_preprocess.py --dataset_path _datasets/kitti/ --seq 0 1 2 3 4 5 6 7 9 10 --out_dir _datasets/kitti/
  3. Generate Ground Truth SDFs (optional for sampling, required for training):

    python scripts/dataset/field_generation/kitti360_field_generation.py --dataset_path _datasets/kitti/ --out_dir _datasets/kitti_fields/ --seq 0 1 2 3 4 5 6 7 9 10 --voxel_size 0.15 --grid_size 96 96 6 --save_files

Waymo Open Dataset

  1. Download the Waymo Open Dataset Perception Dataset v1.4.2 into ./_datasets/waymo/

  2. For preprocessing lidar, bounding boxes, and maps, run the following. Warning: the entire preprocessed Waymo dataset takes around 2TB on top of the raw Waymo tfrecords.

    python scripts/dataset/preprocess/waymo_preprocess.py --tfrecord_paths _datasets/waymo/ --split training validation --out_dir _datasets/waymo_preprocessed/
  3. Generate Ground Truth SDFs:

    python scripts/dataset/field_generation/waymo_field_generation.py --scans _datasets/waymo_preprocessed/ --pose _datasets/waymo_preprocessed/ --splits training validation --out_dir _datasets/waymo_fields/ --voxel_size 0.15 --grid_size 96 96 6 --save_files

3. Assets Generation

We provide a collection of pregenerated assets bank for convenience:

  • Full Version (200+ vehicle meshes, 150+ pedestrian sequences):
    wget -O full_assets.tar.xz "https://uofi.box.com/shared/static/4bndbr8l2fgmjb3tdu1gvjts8hqbo76w.xz"
    tar -xvf full_assets.tar.xz
    rm full_assets.tar.xz
  • Small Version (20 vehicle meshes, 2 pedestrian sequences):
    wget -O small_assets.tar.xz "https://uofi.box.com/shared/static/ir42y4a71luia12u17bk5v0fierqrogw.xz"
    tar -xvf small_assets.tar.xz
    rm small_assets.tar.xz

If you want to generate your own assets, see ASSET_GENERATION.md.

4. Folder Structure

Ultimately, the project directory should look like this

  lidardm/ 
    ├── ...
    ├── _datasets/
    |     ├── waymo/
    |     |     └── ...
    |     ├── waymo_preprocessed/
    |     |     └── ...
    |     ├── waymo_fields/
    |     |     └── ...
    |     ├── kitti/
    |     |     └── ...
    |     └── kitti_fields/
    |           └── ...
    └── generated_assets/
          ├── pedestrian/
          │     └── pedestrian_1/
          │           ├── 000.obj
          │           └── ...
          └── vehicle/
                ├── car1.obj
                └── ...

Training

Our pipeline has 2 learnable components:

  • Scene Generation: We employ a classifier-free guidance-based latent diffusion model to generate the underlying 3D static world. Refer to SCENE_GENERATION.md for details on training.
  • Raydropping: After constructing the 4D world and raycasting on it, we train a raydropping network inspired by RangeNet to make our generated LiDAR more realistic. Refer to RAYDROPPING.md for details on training.

Inference

1. Download Model Weights and Assets

Waymo

  • Prior to receiving access to the Waymo Weights you are required to have a valid Waymo Open Dataset account with access to the Waymo Open Dataset.

  • If you are successfully registered with the Waymo Open dataset, send an email in the following format to receive the Waymo weights.

    • To: vlasz2 AT_SIGN illinois.edu
    • Subject Line: "lidardm waymo weights"
  • You will receive a reply with instructions on getting Map VAE, Scene VAE, Diffusion Model, and raydropping network weights trained on Waymo.

  • You will also receive instructions for setting up the Waymo Baseline Model.

KITTI-360

mkdir pretrained_models && cd pretrained_models

wget -O kitti_weights.zip "https://uofi.box.com/shared/static/tc4hppt38ryy5rsgthiw4q50dxtu1w2f.zip"

unzip kitti_weights.zip

rm kitti_weights.zip

2. Verify Folder Structure

Ultimately, the project directory should look like this

  lidardm/ 
    ├── ...
    └── pretrained_models/
          ├── waymo/ 
          │     ├── raydrop/
          │     │     └── ...
          │     └── scene_gen/
          │           └── ...
          ├── kitti360/ 
          │     ├── raydrop/
          │     │     └── ...
          │     └── scene_gen/
          │           └── ...
          ├── waymo_baseline/
          │     └── ...
          └── waymo_trajbank.npy

3. Run sampling

We have provided the jupyter notebook in examples which provides visualization (like the teaser.gif above), but you can also run larger scale sampling for metrics below.

Sampling

KITTI-360 Model

Sample 2000 samples:

python scripts/diffusion/diffusion_sample_field.py +experiment=kf_s_unet +model.unet.pretrained=../../pretrained_models/kitti360/scene_gen/kfsunet_b.ckpt +sampling.seed_time=True +sampling.outfolder=$PWD/_samples/my_lidardm_kitti +sampling.skiprender=True

Perform raycasting on those samples:

python lidardm/lidar_generation/scene_composition/scripts/unconditional_comp.py --folder _samples/my_lidardm_kitti/ --waymo_dataset_root _datasets/waymo_preprocessed/

The results will be generated under _samples/my_lidardm_kitti. Optionally, you can also run the following script to extract the LiDAR for MMD/JSD calculation.

mkdir _samples/my_lidardm_kitti_npy

python scripts/util/move_lidar.py _samples/my_lidardm_kitti/ _samples/my_lidardm_kitti_npy

Now, the folder _samples/my_lidardm_kitti_npy can be provided to the MMD and JSD evaluation script.

Waymo Model

Random Sampling

The following script samples meshes by randomly acquiring conditions from the dataset, and then runs raycasting:

python scripts/diffusion/diffusion_sample_field.py +experiment=wf_s_unetc +model.unet.pretrained=../../pretrained_models/waymo/scene_gen/wfsunetc.ckpt +sampling.seed_time=True +sampling.outfolder=$PWD/_samples/my_lidardm_waymo +sampling.skiprender=True

python lidardm/lidar_generation/scene_composition/scripts/conditional_comp.py --folder $PWD/_samples/my_lidardm_waymo --waymo_dataset_root _datasets/waymo_preprocessed/

Sampling Dataset

We provide a script to run LidarDM LiDAR generation conditioned on the entire Waymo dataset splits for Sim2Real and Real2Sim evaluation tasks.

For test set generation.

mkdir _datasets/lidardm_waymo_sim_test

python scripts/diffusion/diffusion_generate_dataset.py +experiment=wf_s_unetc +model.unet.pretrained=../../pretrained_models/waymo/scene_gen/wfsunetc.ckpt +sampling.outfolder=$PWD/_datasets/lidardm_waymo_sim_test +sampling.skiprender=True +sampling.use_test=True

python lidardm/lidar_generation/scene_composition/scripts/conditional_comp.py --folder $PWD/_datasets/lidardm_waymo_sim_test --waymo_dataset_root _datasets/waymo_preprocessed/

For train set generation.

mkdir _datasets/lidardm_waymo_sim_train

python scripts/diffusion/diffusion_generate_dataset.py +experiment=wf_s_unetc +model.unet.pretrained=../../pretrained_models/waymo/scene_gen/wfsunetc.ckpt +sampling.outfolder=$PWD/_datasets/lidardm_waymo_sim_train +sampling.skiprender=True

python lidardm/lidar_generation/scene_composition/scripts/conditional_comp.py --folder $PWD/_datasets/lidardm_waymo_sim_train --waymo_dataset_root _datasets/waymo_preprocessed/

Evaluating Metrics

MMD/JSD

The MMD/JSD evaluation script is designed to compare two folders of point clouds (in the format folder1/*.npy). We provide a pregenerated set of samples used for comparison in the paper.

mkdir _samples && cd _samples

wget -O kitti_uncond.zip https://uofi.box.com/shared/static/ulommpmbq064azfjedkrgzlbmd7drwzv.zip

unzip kitti_uncond.zip

The following command executes MMD calculation.

python scripts/metric/eval_set.py --folder1 _samples/kitti_uncond/kitti/ --folder2 _samples/kitti_uncond/lidardm/ --type mmd --folder2_rotations 0

To perform JSD calculation, replace --type mmd with --type jsd

Point2Plane and Chamfer

For a folder of LidarDM generation on Waymo run:

python scripts/metric/point2plane.py --input_folder _samples/my_lidardm_waymo/ 

To generate samples for the voxel sequence diffusion model baseline, see the instructions in VOXEL_BASELINE.md.

To evaluate the results from the baseline model run:

python scripts/metric/point2plane.py --is_baseline True --input_folder _samples/my_voxelseq_waymo/

Planner

For instructions on training the LiDAR planner and running planner evaluations refer to PLANNER.md