A comprehensive testing script for VGGT (Visual Geometry Grounded Transformer) with multi-view 3D reconstruction capabilities.
VGGT is a state-of-the-art neural network for multi-view 3D reconstruction presented at CVPR 2025 by Oxford and Meta providing:
- Camera pose estimation (direct output, no PnP needed)
- Depth maps with confidence scores
- 3D point clouds (both from depth and direct 3D prediction)
- COLMAP format export for integration with other tools
- Comprehensive timing analysis
# Clone VGGT repository
git clone https://github.com/facebookresearch/vggt.git
cd vggtFollow instrunctions found at VGGT GitHub Repository
The VGGT-1B model (5GB) will be automatically downloaded from Hugging Face on first use and cached locally. Subsequent runs use the cached model.
python vggt_experimenter.py \
--images /path/to/scene/images \
--workspace /path/to/output \
--conf_threshold 65.0python vggt_experimenter.py \
--images /path/to/scene/images \
--workspace /path/to/output \
--size 512 \
--dtype fp16 \
--conf_threshold 65.0 \
--model_size 1Bpython vggt_experimenter.py \
--dataset_root /path/to/multiple/scenes \
--workspace /path/to/output \
--conf_threshold 65.0| Parameter | Description | Default | Recommended Values |
|---|---|---|---|
--images |
Path to folder containing scene images | Required | - |
--dataset_root |
Path to parent folder for batch processing | Alternative to --images | - |
--workspace |
Output directory for results | Required | - |
--size |
Resize longer edge (pixels) | 512 | 224-512 (smaller = less memory) |
--dtype |
Inference precision | fp16 | fp16, bf16, fp32 |
--conf_threshold |
Confidence percentile (0-100%) | 50.0 | See guidelines below |
--model_size |
VGGT model variant | 1B | Currently only 1B available |
The confidence threshold is specified as a percentile (0-100%), where higher values keep only the most confident points:
| Scene Type | Confidence Threshold | Reasoning |
|---|---|---|
| Large outdoor scenes (Tanks & Temples) | 65-80% | Complex geometry, occlusions, varying lighting |
| Object-centric scenes | 0-50% | Simpler geometry, controlled conditions |
| Indoor scenes | 10-65% | Moderate complexity, some occlusions |
# Large outdoor scene (Tanks & Temples style)
--conf_threshold 75.0
# Simple object reconstruction
--conf_threshold 45.0
# Mixed indoor/outdoor scene
--conf_threshold 60.0The script generates comprehensive outputs in the workspace directory:
workspace/
├── cameras_extrinsic.npy # Camera poses [S, 4, 4]
├── cameras_intrinsic.npy # Camera intrinsics [S, 3, 3]
├── pointcloud_from_depth.ply # Point cloud from depth maps
├── pointcloud_from_points.ply # Point cloud from direct 3D prediction
├── timings.csv # Performance timing breakdown
├── colmap/ # COLMAP format export
│ ├── cameras.txt
│ ├── images.txt
│ └── points3D.bin
└── depth_maps/ # Depth and confidence visualizations
├── depth_000.png
├── conf_000.png
└── ...
VGGT-1B requires significant GPU memory. For RTX 4090 (24GB):
| Image Count | Resolution | Memory Usage | Recommendation |
|---|---|---|---|
| 10-20 images | 512x512 | ~8-12GB | Safe |
| 50 images | 224x224 | ~12-16GB | Works |
| 100+ images | 224x224 | >24GB | OOM on RTX 4090 |
Note: On RTX 4090, you will get OOM with 100 images at 224x224 resolution.