DepR: Depth Guided Single-view Scene Reconstruction with Instance-level Diffusion

Qingcheng Zhao^*,1,† · Xiang Zhang^*,✉,2 · Haiyang Xu² · Zeyuan Chen² · Jianwen Xie³ · Yuan Gao⁴ · Zhuowen Tu²

¹ShanghaiTech University · ²UC San Diego · ³Lambda, Inc. · ⁴Stanford University

ICCV 2025

^* equal contribution    ^✉ corresponding author

^† Project done while Qingcheng Zhao interned at UC San Diego.

Project Page | Paper | arXiv

🛠️ Environment Setup

We provide a pre-built Docker image at zx1239856/depr based on PyTorch 2.7.1 and CUDA 12.6. You can also build the image locally:

docker build -f Dockerfile . -t depr

Alternatively, you can install dependencies based on commands listed in Dockerfile.

🗂️ Dataset Setup

Please download processed 3D-FRONT dataset from https://huggingface.co/datasets/zx1239856/DepR-3D-FRONT. Extract the downloaded files into datasets/front3d_pifu/data. The result folder structure should look like

data/
|-- metadata/             (Scene metadata)
|   |-- 0.jsonl
|   |-- ...
|-- pickled_data/         (Raw data processed by InstPIFu)
|   |-- test/
|       |-- rendertask3000.pkl
|       |-- ...
|-- sdf_layout/           (GT layouts)
|   |-- 10000.npy
|   |-- ...
|- 3D-FUTURE-watertight/  (GT meshes, required for evaluation)
|   |-- 0004ae9a-1d27-4dbd-8416-879e9de1de8d/
|       |-- raw_watertight.obj
|       |-- ...
|-- instpifu_mask/        (Instance masks provided by InstPIFu)
|-- panoptic/             (Panoptic segmentation maps we rendered)
|-- img/                  (Optional, can be extracted from pickled data)
|-- depth/depth_pro/      (Optional)
`-- grounded_sam/         (Optional)

Alternatively, you may generate depth / segmentation yourself based on instructions below.

Generate Segmentation

Please prepare Grounded SAM weights in checkpoint/grounded_sam.

grounded_sam/
|-- GroundingDINO_SwinB.py
|-- groundingdino_swinb_cogcoor.pth
|-- groundingdino_swint_ogc.pth
`-- sam_vit_h_4b8939.pth

python -m scripts.run_grounded_sam

Generate Depth

Please put Depth Pro weights in checkpoint/.

python -m scripts.run_depth_pro --output depth_pro

📊 Inference

Please download our weights from https://huggingface.co/zx1239856/DepR and put everything in the checkpoint folder.

🚀 Demo

We provide a demo.ipynb notebook for inference demo on real-world images.

Object-level Evaluation

You may change 8 to the actual number of GPUs as needed.

bash launch.sh 8 all

(Optional) Guided Sampling

bash launch.sh 8 all --guided

Scene-level Evaluation

# Generate shapes
bash launch.sh 8 sample --metadata datasets/front3d_pifu/meta/test_scene.jsonl --use-sam

# Layout optim
bash launch.sh 8 scene --use-sam

# Prepare GT scene
python -m scripts.build_gt --out-dir output/gt

# Calculate scene-level CD/F1
accelerate launch --num_processes=8 --multi_gpu -m scripts.eval_scene --gt-pcd-dir output/gt/pcds --pred-dir output/infer/sam_3dproj_attn_dino_c9_augdep_augmask_nocfg_model_0074999/ --save-dir output/evaluation/results --method depr

🏷️ License

This repository is released under the CC-BY-SA 4.0 license.

🙏 Acknowledgement

Our framework utilizes pre-trained models including Grounded-Segment-Anything, Depth Pro, and DINO v2.

Our code is built upon diffusers, Uni-3D, and BlockFusion.

We use physically based renderings of 3D-FRONT scenes provided by InstPIFu. Additionally, we rendered panoptic segmentation maps ourselves.

We thank all these authors for their nicely open sourced code/datasets and their great contributions to the community.

📝 Citation

If you find our work useful, please consider citing:

@InProceedings{Zhao_2025_ICCV_DepR,
    author    = {Zhao, Qingcheng and Zhang, Xiang and Xu, Haiyang and Chen, Zeyuan and Xie, Jianwen and Gao, Yuan and Tu, Zhuowen},
    title     = {DepR: Depth Guided Single-view Scene Reconstruction with Instance-level Diffusion},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {5722-5733}
}