InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes

Project Page | Paper | Arxiv | Video

InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes,
Zesong Yang, Bangbang Yang, Wenqi Dong, Chenxuan Cao, Liyuan Cui, Yuewen Ma, Zhaopeng Cui, Hujun Bao
ICCV 2025

teaser.mp4

Installation

• Installation of Scene Decomposition.

conda create -n instascene python=3.9 -y
conda activate instascene 

pip install torch==2.1.0+cu118 torchvision==0.16.0+cu118 torchaudio==2.1.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

pip install --extra-index-url=https://pypi.nvidia.com "cudf-cu11==24.2.*" "cuml-cu11==24.2.*"

pip install -r requirements.txt

Install CropFormer for instance-level segmentation.

cd semantic_modules/CropFormer
cd mask2former/modeling/pixel_decoder/ops
sh make.sh
cd ../../../../
git clone [email protected]:facebookresearch/detectron2.git
cd detectron2
pip install -e .
pip install git+https://github.com/cocodataset/panopticapi.git
pip install git+https://github.com/mcordts/cityscapesScripts.git
cd ..
pip install -r requirements.txt
pip install -U openmim
mim install mmcv
mkdir ckpts

Manually download CropFormer checkpoint into semantic_modules/CropFormer/ckpts

• Installation of in-situ generation.

Data Preprocessing

Please follow the steps below to process your custom dataset, or directly download our preprocessed datasets.

1. Run instance-level segmentation.

• It's ok to use other 2D segmentation models, but make sure the input masks don't exhibit overly complex hierarchy relationships; otherwise, our method will default to the finest level.

cd semantic_modules/CropFormer
bash run_segmentation.sh "$DATA_DIR"
cd ../..

2. Training 2DGS.

Follow the original repository to train the 2dgs model.

python train.py -s data/3dovs/bed -m output/3dovs/bed/train_2dgs

Optional mono normal prior (StableNormal) is available to enhance the reconstruction quality.

## Prepare Normal Priors
cd semantic_modules
git clone https://github.com/Stable-X/StableNormal && cd StableNormal
pip install -r requirements.txt
mv ../inference_stablenormal.py ./
python inference_stablenormal.py "$DATA_DIR"
cd ../..

## Training 2DGS with Normal Priors 
python train.py -s data/3dovs/bed --w_normal_prior stablenormal_normals -m output/3dovs/bed/train_2dgs

Put the trained point_cloud.ply file into the $DATA_DIR directory. After successfully executing the above steps, the data directory should be structured as follows:

data
   |——————3D_OVS
   |   |——————bed
   |      |——————point_cloud.ply
   |      |——————images
   |         |——————00.jpg
   |         ...
   |      |——————sam
   |         |——————mask
   |            |——————00.png
   |            ...
   |      |——————sparse
   |         |——————0
   |            |——————cameras.bin
   |            ...
   |      |——————(optional) stablenormal_normals
   |         |——————00.png
   |         ...
   |     ...

Training with Spatial Contrastive Learning

Note that for simple scenes, such as 3D-OVS (simple-object centered without overlap), no need to use spatial relationships to obtain robust semantic priors as shown in our supplementary material. Single-view constrastive learning is sufficient to achieve strong performance.

We train the model on a NVIDIA Tesla A100 GPU (40GB) with 10,000 iterations for about 20 minutes & less than 8GB GPU.

• Reduce the GPU & Speed the time with --sample_batchsize 8 * 1024 or -r 2.
• Use --gram_feat_3d for a more robust feature field in complex scenes.
• It's normal to get stuck at the DBScan Filter Stage, since the backgrount gaussian points may be divided into multi-regions.
• Use --consider_negative_labels to suppress floaters during background segmentation.

python train_semantic.py -s data/lerf/waldo_kitchen \
                         -m train_semanticgs \
                         --use_seg_feature --iterations 10000 \
                         --load_filter_segmap --consider_negative_labels

After completing the training, we provide a GUI modified from Omniseg3D for real-time ineractive segmentation. The point_cloud.ply in our preprocessed datasets already has pretrained semantic features.

python semantic_gui.py \
  --ply_path data/lerf/waldo_kitchen/point_cloud.ply \
  --interactive_note lerf_waldo_kitchen \
  --use_colmap_camera \
  --source_path data/lerf/waldo_kitchen --resolution 1

• Left Mouse for changing rendering view
• Click Mode + 0.9 Threshold + Right Mouse for segmentation
• Clear Edit for clear the segmentation cache
• Delete 3D for remove the chosen gaussians
• Segment 3D for only keep the chosen gaussians
• Reload Data for reload the gaussian model

Screencast.2025-07-24.13_31_27.mp4

Feishu20250723-192829.mp4

ToDos

🔥 Feel free to raise any requests, including support for additional datasets or broader applications of segmentation~

• Release project page and paper.
• Release scene decomposition code.
• Release in-situ generation code.

Acknowledgements

Some codes are modified from Omniseg3D, MaskClustering, 2DGS++, thanks for the authors for their valuable works.

Citation

If you find this code useful for your research, please use the following BibTeX entry.

@inproceedings{yang2025instascene,
    title={InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes},
    author={Yang, Zesong and Yang, Bangbang and Dong, Wenqi and Cao, Chenxuan and Cui, Liyuan and Ma, Yuewen and Cui, Zhaopeng and Bao, Hujun},
    booktitle=ICCV,
    year={2025}
}