VGGT-CD: Training-Free Robust Registration for 3D Change Detection

Paper | Code

Official implementation of VGGT-CD, a training-free coarse-to-fine pipeline for 3D change detection from unposed multi-view images.

Abstract: Independent reconstructions of two time epochs suffer from scale ambiguity, the registration-change paradox, and edge-flying noise. VGGT-CD uses sparse joint inference to establish a shared metric space and an initial Sim(3) prior, then refines dense reconstructions through reliability-guided purification and closed-form translation alignment. On the 11-scene World Across Time benchmark, it reduces absolute trajectory error by 44% outdoors and 59% indoors while completing registration more than 6x faster than prior baselines.

Figure 1. VGGT-CD aligns independent bi-temporal reconstructions and isolates genuine 3D changes, resolving scale ambiguity, dynamic outliers, and edge noise.

Method

VGGT-CD addresses the coordinate and scale inconsistency between independently reconstructed T1/T2 image sequences.

1. VGGT reconstructs dense geometry, depth, confidence, and camera poses for T1 and T2.
2. A sparse joint-inference stage estimates a cross-time Sim(3) transform in a unified metric space.
3. High-confidence dense points are filtered to remove dynamic changes and edge noise.
4. Static correspondences are refined with voxel hashing and one-shot SVD.
5. The final alignment is saved as fine_sim3.npz.

Repository Structure

VGGT-CD/
|- run_inference.py               # Main VGGT-CD pipeline
|- run_inference.sh               # Shell wrapper for inference
|- coarse_to_fine_registration.py # Core coarse-to-fine registration method
|- IMPROVEMENTS.md                # Method details
|- app.py                         # Gradio demo entry
|- vggt/                          # VGGT model code
|- experiments/
|  |- baselines/                 # Baseline methods
|  |- evaluation/                # Evaluation scripts
|  `- results/summary.csv        # Experiment summary
|- docs/                          # GitHub Pages project site
`- requirements.txt

Installation

git clone https://github.com/WZ-CS/VGGT-CD.git
cd VGGT-CD

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Set the VGGT checkpoint path:

export VGGT_MODEL_PATH=/path/to/model.pt

Data Format

DATA_ROOT/
`- scene_name/
    |- images/
    |  |- SEQ_T1/
    |  `- SEQ_T2/
    `- sparse/0/
        |- cameras.bin
        |- images.bin
        `- points3D.bin

Inference

python run_inference.py \
  --data_root /path/to/WAT \
  --scene breville \
  --seq_t1 IMG_9184 \
  --seq_t2 IMG_9185 \
  --output_root ./eval_results \
  --model_path "${VGGT_MODEL_PATH}" \
  --num_keyframes 5 \
  --conf_threshold 50

Shell wrapper:

DATA_ROOT=/path/to/WAT \
SCENE=breville \
SEQ_T1=IMG_9184 \
SEQ_T2=IMG_9185 \
VGGT_MODEL_PATH=/path/to/model.pt \
./run_inference.sh

Evaluation

python experiments/evaluation/eval_poses_1.py --eval_root ./eval_results

Scene outputs:

eval_results/scene_name/
|- pred_poses_t1.npy
|- pred_poses_t2.npy
|- gt_poses_t1.npy
|- gt_poses_t2.npy
|- image_names_t1.txt
|- image_names_t2.txt
|- coarse_sim3.npz
|- fine_sim3.npz
`- timing.json

Results

Scenes	ATE mean	ATE RMSE	RTE mean	RRE mean	Time/scene	GPU memory
11	0.4316	0.4631	0.1933	0.3598	33.14 s	18167.52 MB

Baselines

Script	Method
experiments/baselines/run_baseline_vggt_only.py	Independent VGGT T1/T2 inference
experiments/baselines/baseline_icp.py	VGGT point clouds with ICP
experiments/baselines/run_baseline_fgr.py	VGGT point clouds with Fast Global Registration
experiments/baselines/run_baseline_global_scale_icp.py	Global scale alignment with ICP
experiments/baselines/run_baseline_sim3_icp.py	Sim(3) alignment with ICP
experiments/baselines/run_baseline_geotransformer.py	GeoTransformer baseline

Demo

GRADIO_SERVER_PORT=7860 python app.py