Coming Soon

SCaR-3D: 3D Scene Change Modeling With Consistent Multi-View Aggregation
3DV 2026

Zirui Zhou¹, Junfeng Ni^1,2, Shujie Zhang¹, Yixin Chen^2✉, Siyuan Huang^2✉

¹Tsinghua University    ²State Key Laboratory of General Artificial Intelligence, BIGAI

SCAR-3D is a novel 3D scene change detection and reconstruction framework that identifies object-level changes from dense pre-change and sparse post-view images. It leverages a signed-distance-based 2D differencing module, multi-view aggregation with voting and pruning, and segmentation validation to produce accurate and consistent 3D change masks. The method also supports continual scene reconstruction by selectively updating dynamic regions.

---

🌟Features

• Multi-view Consistent Change Detection: Aggregates 2D differences into a unified 3D representation with voting and pruning.
• Signed-Distance-Based Localization: Captures directional changes in feature space for robust detection.
• Segmentation Validation: Uses EfficientSAM to refine change masks and improve accuracy.
• Continual Reconstruction: Updates only changed regions to maintain scene consistency and reduce artifacts.
• Synthetic Dataset (CCS3D): Provides editable indoor scenes with diverse change types for controlled evaluation.

---

🔨 Installation

Prerequisites

• CUDA 12.4 or higher
• Python 3.8
• Conda (recommended)

Step 1: Clone the Repository

git clone https://github.com/zr-zhou0o0/SCaR-3D.git
cd SCaR-3D

Step 2: Create Conda Environment

conda create -n scar3d python=3.8 -y
conda activate scar3d

Step 3: Install PyTorch and CUDA Dependencies

pip install torch==2.4.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

Step 4: Install Other Dependencies

pip install -r requirements.txt

Step 5: Install Other Required Packages from Source

pip install git+https://github.com/yformer/EfficientSAM.git@c9408a74b1db85e7831977c66e9462c6f4891729

pip install git+https://github.com/camenduru/simple-knn.git

pip install git+https://github.com/rahul-goel/fused-ssim.git

pip install submodules/diff-point-rasterization

📦 Pretrained Weights

Download the pretrained EfficientSAM weights here.

Download the pretrained CCS3D model weights here.

Download the pretrained 3DGS-CD model weights here.

---

📊 Dataset

CCS3D (Controllable Change in 3D Scenes)

A synthetic dataset built with Blender, featuring four complex indoor scenes:

• Desk
• Bookcase
• Livingroom
• Bedroom

Each scene supports:

• Multiple change types: insertion, removal, translation, rotation, mixed.
• Complex camera trajectories simulating real-world navigation.
• Fine-grained object-level annotations.

Download the CCS3D dataset here.

3DGS-CD

Our processed 3DGS-CD dataset can be downloaded here.

Customized Datasets

Dataset Structure

The dataset should be organized as follows:

dataset/
└── <dataset_name>/
    ├── <scene_name>/
    │   ├── images/          # All images
    │   ├── train-pre/       # Pre-change training images
    │   ├── train-post/      # Post-change training images
    │   ├── test-pre/        # Pre-change test images
    │   ├── test-post/       # Post-change test images
    │   ├── gt-pre-mask/     # Ground truth masks for pre-change
    │   ├── gt-post-mask/    # Ground truth masks for post-change
    │   ├── sparse/          # COLMAP sparse reconstruction
    │   └── <scene_name>.db  # COLMAP database
    └── ...

---

💡Usage

Change Detection

# CCS3D Dataset
bash run_ccs3d.sh

# 3DGS-CD Dataset
bash run_3dgs_cd.sh

# If you want to train 3DGS models from scratch, use the following command:
bash run_train.sh

---

Results

Quantitative Change Detection Results

Results on CCS3D Dataset

Method	Livingroom F1	Livingroom IoU	Desk F1	Desk IoU	Bookcase F1	Bookcase IoU	Bedroom F1	Bedroom IoU	Average F1	Average IoU
Pixel-Diff	0.273	0.162	0.398	0.254	0.315	0.201	0.286	0.176	0.318	0.198
Feature-Diff	0.420	0.302	0.480	0.323	0.320	0.256	0.705	0.584	0.450	0.343
CL-Splat	0.789	0.657	0.567	0.399	0.294	0.199	0.501	0.341	0.538	0.399
MV3DCD	0.478	0.329	0.291	0.178	0.449	0.295	0.547	0.413	0.441	0.304
3DGS-CD	0.897	0.815	0.525	0.408	0.477	0.353	0.148	0.089	0.512	0.416
Ours	0.955	0.914	0.610	0.477	0.423	0.377	0.909	0.834	0.724	0.650

Qualitative Examples

---

Citation

@inproceedings{zhou2026scar3d,
  title={3D Scene Change Modeling With Consistent Multi-View Aggregation},
  author={Zhou, Zirui and Ni, Junfeng and Zhang, Shujie and Chen, Yixin and Huang, Siyuan},
  booktitle=ThreeDV,
  year={2026}
}

---