Nerfstudio Trainer

Overview

This skill launches nerfstudio splatfacto (3D Gaussian Splatting) training with parameters tuned for LiDAR-camera fusion datasets. It manages the full launch workflow: parameter configuration, background execution, and log file setup.

When to Use This Skill

Use this skill when the user requests to:

• "学習を開始して" / "trainを走らせて"
• "新しいパラメータで再学習"
• "splatfactoを起動"
• "3DGS training を開始"
• "v{N}で学習開始"

How to Use This Skill

Step 1: Determine Parameters

Ask or infer the following from user context:

Step 2: Determine Log Version

Check existing log files to auto-increment:

ls {project_dir}/train_lidar_v*.log 2>/dev/null | sort -V | tail -1

Step 3: Build and Execute Command

Base Command Template

ns-train splatfacto \
  --data {data_dir} \
  --output-dir {output_dir} \
  --max-num-iterations {max_num_iterations} \
  --pipeline.model.camera-optimizer.mode {camera_optimizer} \
  --pipeline.model.reset-alpha-every {reset_alpha_every} \
  --pipeline.model.densify-grad-thresh {densify_grad_thresh} \
  --pipeline.datamanager.dataparser.orientation-method none \
  --pipeline.datamanager.dataparser.center-method poses \
  --pipeline.datamanager.dataparser.auto-scale-poses False \
  --pipeline.datamanager.dataparser.assume-colmap-world-coordinate-convention False \
  --pipeline.datamanager.dataparser.load-3D-points {load_3d_points} \
  > {project_dir}/train_lidar_v{N}.log 2>&1 &

Important Notes on Parameters

Densification fix for large datasets (>3000 images):

• nerfstudio has a known issue (GitHub #2927, #3743) where pause_refine_after_reset = num_train_data + refine_every
• Default reset_alpha_every=30 → reset_every=3000, which is smaller than pause_refine_after_reset for large datasets
• Must use reset_alpha_every=200 (or higher) so that reset_every=20000 > pause_refine_after_reset
• Without this fix, Gaussian count stays constant (no densification)

LiDAR-based datasets:

• orientation-method none: LiDAR world coordinates are already aligned
• center-method poses: Center scene at camera pose centroid
• auto-scale-poses False: Preserve real-world scale from LiDAR SLAM
• assume-colmap-world-coordinate-convention False: Using LiDAR world frame, not COLMAP convention

Step 4: Verify Launch

After launching:

1. Confirm PID is running: ps aux | grep ns-train
2. Check initial log output: tail -5 {log_file}
3. Report the viewer URL (typically http://localhost:7007)
4. Report the output directory path

Step 5: Report to User

Provide:

• PID
• Log file path
• Output directory
• Viewer URL
• Key parameters used
• Expected densification window (when Gaussians should start growing)

Parameter Presets

Default (LiDAR + large dataset)

reset_alpha_every=200
densify_grad_thresh=0.0002
camera_optimizer=SO3xR3
max_num_iterations=30000

Fast test (quick validation)

reset_alpha_every=200
densify_grad_thresh=0.0002
camera_optimizer=SO3xR3
max_num_iterations=5000

High quality (longer training)

reset_alpha_every=200
densify_grad_thresh=0.0001
camera_optimizer=SO3xR3
max_num_iterations=50000

Troubleshooting

NaN error at startup:

• Check points3D.txt for NaN/Inf values
• Filter with: np.isfinite(xyz).all(axis=1)

Gaussian count not growing:

• Verify reset_alpha_every is large enough: reset_alpha_every × refine_every > num_train_data + refine_every
• Check if training step has passed pause_refine_after_reset

CUDA out of memory:

• Reduce initial point count in points3D.txt
• Lower max-num-iterations for testing