Train on an NCore v4 Capture

The examples/simple_trainer.py script supports training a 3D Gaussian Splatting model on an NCore v4 capture. Point --data-dir at the sequence meta-JSON file and set --data-type ncore:

cd examples
python simple_trainer.py default \
    --data-type ncore \
    --data-dir /path/to/sequence.json \
    --data-factor 1 \
    --result-dir results/my_scene \
    --init-type lidar

For datasets with FTheta cameras (auto-detected from sequence metadata):

cd examples
python simple_trainer.py mcmc \
    --data-type ncore \
    --data-dir /path/to/sequence.json \
    --data-factor 1 \
    --result-dir results/my_scene \
    --init-type lidar \
    --with-ut \
    --with-eval3d

Note

The camera model (pinhole, fisheye, or FTheta) is read automatically from the NCore sequence metadata. Passing --camera-model has no effect when --data-type ncore is used.

Note

Training with --data-factor values other than 1 requires that all camera resolutions are divisible by the factor.

Note

World-space normalization (PCA + scale) is applied by default (--normalize-world-space is True). For outdoor or large-scale captures where preserving real-world scale matters, pass --no-normalize-world-space.

Note

When a sequence contains more than one camera or lidar sensor, you must explicitly specify which sensor(s) to use via --ncore-camera-ids and/or --ncore-lidar-ids. If omitted, the parser auto-detects sensors but will raise an error when multiple sensors are present to avoid ambiguity (e.g. when multiple downscaled variants of the same sensor exist).

cd examples
python simple_trainer.py default \
    --data-type ncore \
    --data-dir /path/to/sequence.json \
    --ncore-camera-ids camera_front \
    --ncore-lidar-ids lidar_top \
    ...