SUDS: Scalable Urban Dynamic Scenes

Haithem Turki, Jason Y. Zhang, Francesco Ferroni, Deva Ramanan

This repository contains the code needed to train SUDS models.

Citation

@misc{turki2023suds,
   title={SUDS: Scalable Urban Dynamic Scenes},
   author={Haithem Turki and Jason Y. Zhang and Francesco Ferroni and Deva Ramanan},
   year={2023},
   eprint={2303.14536},
   archivePrefix={arXiv},
   primaryClass={cs.CV}
}

Setup

conda env create -f environment.yml
conda activate suds
python setup.py install

The codebase has been mainly tested against CUDA >= 11.3 and A100/A6000 GPUs. GPUs with compute capability greater or equal to 7.5 should generally work, although you may need to adjust batch sizes to fit within GPU memory constraints.

Data Preparation

KITTI

Download the following from the KITTI MOT dataset:
1. Left color images
2. Right color images
3. GPS/IMU data
4. Camera calibration files
5. Velodyne point clouds
6. (Optional) Semantic labels
Extract everything to ./data/kitti and keep the data structure
Generate depth maps from the Velodyne point clouds: python scripts/create_kitti_depth_maps.py --kitti_sequence $SEQUENCE
(Optional) Generate sky and static masks from semantic labels: python scripts/create_kitti_masks.py --kitti_sequence $SEQUENCE
Create metadata file: python scripts/create_kitti_metadata.py --config_file scripts/configs/$CONFIG_FILE
Extract DINO features:
1. python scripts/extract_dino_features.py --metadata_path $METADATA_PATH or python -m torch.distributed.run --standalone --nnodes=1 --nproc_per_node $NUM_GPUS scripts/extract_dino_features.py --metadata_path $METADATA_PATH for multi-GPU extraction
2. python scripts/run_pca.py --metadata_path $METADATA_PATH
Extract DINO correspondences: python scripts/extract_dino_correspondences.py --metadata_path $METADATA_PATH or python -m torch.distributed.run --standalone --nnodes=1 --nproc_per_node $NUM_GPUS scripts/extract_dino_correspondences.py --metadata_path $METADATA_PATH for multi-GPU extraction
(Optional) Generate feature clusters for visualization: python scripts/create_kitti_feature_clusters.py --metadata_path $METADATA_PATH --output_path $OUTPUT_PATH

VKITTI2

Download the following from the VKITTI2 dataset:
1. RGB images
2. Depth images
3. Camera intrinsics/extrinsics
4. (Optional) Ground truth forward flow
5. (Optional) Ground truth backward flow
6. (Optional) Semantic labels
Extract everything to ./data/vkitti2 and keep the data structure
(Optional) Generate sky and static masks from semantic labels: python scripts/create_vkitti2_masks.py --vkitti2_path $SCENE_PATH
Create metadata file: python scripts/create_vkitti2_metadata.py --config_file scripts/configs/$CONFIG_FILE
Extract DINO features:
1. python scripts/extract_dino_features.py --metadata_path $METADATA_PATH or python -m torch.distributed.run --standalone --nnodes=1 --nproc_per_node $NUM_GPUS scripts/extract_dino_features.py --metadata_path $METADATA_PATH for multi-GPU extraction
2. python scripts/run_pca.py --metadata_path $METADATA_PATH
If not using the ground truth flow provided by VKITTI2, extract DINO correspondences: python scripts/extract_dino_correspondences.py --metadata_path $METADATA_PATH or python -m torch.distributed.run --standalone --nnodes=1 --nproc_per_node $NUM_GPUS scripts/extract_dino_correspondences.py --metadata_path $METADATA_PATH for multi-GPU extraction
(Optional) Generate feature clusters for visualization: python scripts/create_vkitti2_feature_clusters.py --metadata_path $METADATA_PATH --vkitti2_path $SCENE_PATH --output_path $OUTPUT_PATH

Training

python suds/train.py suds --experiment-name $EXPERIMENT_NAME --pipeline.datamanager.dataparser.metadata_path $METADATA_PATH [--pipeline.feature_clusters $FEATURE_CLUSTERS]

Evaluation

python suds/eval.py --load_config $SAVED_MODEL_PATH or python -m torch.distributed.run --standalone --nnodes=1 --nproc_per_node $NUM_GPUS suds/eval.py --load_config $SAVED_MODEL_PATH for multi-GPU evaluation

Acknowledgements

This project is built on Nerfstudio and tiny-cuda-nn. The DINO feature extraction scripts are based on ShirAmir's implementation and parts of the KITTI processing code from Neural Scene Graphs.

hturki/suds