Yang Gao, Wuyang Li, Po-Chien Luan, Alexandre Alahi
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
configs/ Release training and evaluation config
groundedsam/ GroundingDINO + SAM semantic pseudo-label generation
mmdet3d/ DeGO model, dataset, pipeline, and hook code
requirements/ Runtime dependency pins
scripts/ Setup, pseudo-label, train, and eval entry points
tools/ Training, evaluation, VGGT cache, and nuScenes info tools
vggt/ Local VGGT implementation used by cache generation
The validated CUDA image is:
nvidia/cuda:11.3.1-cudnn8-devel-ubuntu20.04
Create the CUDA 11.3 conda environment:
./scripts/setup_env_cu113.shActivate it manually:
source /path/to/miniconda3/etc/profile.d/conda.sh
conda activate dego_cu113The repo expects nuScenes, occupancy ground truth, pseudo labels, dataset info
files, and the VGGT cache under data/.
Expected paths:
data/nuscenes
data/gts
data/grounded_sam_nusc
data/metric_3d_nusc
data/vggt_cache_spatial_temporal_block22
data/bevdetv2-nuscenes_infos_train.pkl
data/bevdetv2-nuscenes_infos_val.pkl
To regenerate nuScenes info files when needed:
python tools/create_data_bevdet.py \
--root-path data/nuscenes \
--out-dir data \
--version v1.0-trainvalWe follow GaussianFlowOcc to process Nuscenes data and generate pseudo depth/segmentation labels.
Generate Metric3D depth pseudo labels:
./scripts/generate_metric3d_depth.shGenerate GroundingDINO + SAM semantic masks:
SAM_CHECKPOINT=ckpts/sam_vit_h_4b8939.pth \
./scripts/generate_groundedsam_masks.shGenerate the VGGT block-22 spatial-temporal feature cache:
./scripts/generate_vggt_cache.shDefault four-GPU release training:
./scripts/train_release_4gpu.shUseful overrides:
CONFIG=configs/dego_30e_vggt_after20.py \
WORK_DIR=work_dirs/dego_30e_vggt_after20_release \
GPUS=4 PORT=29500 AUTO_RESUME=1 \
./scripts/train_release_4gpu.shEquivalent direct command:
PORT=29500 bash tools/dist_train.sh \
configs/dego_30e_vggt_after20.py \
4 \
--work-dir work_dirs/dego_30e_vggt_after20_release \
--auto-resumeEvaluate one checkpoint:
CHECKPOINT=work_dirs/dego_30e_vggt_after20_release/epoch_30_ema.pth \
./scripts/eval_release.shEquivalent direct command:
python tools/test.py \
configs/dego_30e_vggt_after20.py \
work_dirs/dego_30e_vggt_after20_release/epoch_30_ema.pth \
--eval mIoU \
--nbh 5This work builds on the following projects:
@inproceedings{gao2026deformable,
title={Deformable Gaussian Occupancy: Decoupling Rigid and Nonrigid Motion with Factorized Distillation},
author={Gao, Yang and Li, Wuyang and Luan, Po-Chien and Alahi, Alexandre},
booktitle={The IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026},
year={2026}
}