InstantSplat Test Scene Results

This dataset contains the reconstruction results from InstantSplat on a test scene with 10 RGB images.

Files

• point_cloud_test_scene.ply: 3D Gaussian Splatting point cloud (462MB, 1.86M points)
• test_scene/input_images/: 10 input RGB images (634x475 pixels)
• test_scene/cameras.json: Camera parameters estimated by MASt3R
• test_scene/cfg_args: Training configuration
• test_scene/train_time.txt: Training time statistics

Training Details

• Method: InstantSplat (sparse-view 3D Gaussian Splatting)
• Input: 10 RGB images from ThermoPhysical_3DGS dataset
• Training iterations: 1000
• Training time: ~56 seconds
• Initial points: 1,865,762
• Features:
  • Automatic camera pose optimization
  • No COLMAP preprocessing required
  • Direct 3D reconstruction from images using MASt3R

Performance

• Geometry initialization: 37 seconds
• Joint optimization: 56 seconds
• Total time: < 2 minutes
• Training loss: 0.259 → 0.153

Source

Generated using InstantSplat: https://github.com/NVlabs/InstantSplat