Skip to content
360 Gaussian Pro logo360 Gaussian Pro

Masking

Remove the camera operator, moving objects and any unwanted elements/objects.

Masking removes things you don't want in the reconstruction — the person holding the camera, moving objects, the sky — so they don't confuse alignment or leave ghosts in the final splat. For the short version, see the First Run walkthrough.

Masking runs on NVIDIA, AMD and Intel GPUs. On NVIDIA it uses CUDA; on AMD and Intel it runs the same SAM 3 model through DirectML, about 3× slower but with the same tools, object picker included.

The masking step downloads an AI model the first time you use it — about 1.9 GB on NVIDIA, 3.7 GB on AMD and Intel. This is a one-time download.

The three tools

SAM3 (keyword masks)

Type what you want gone — person, car, sky — and the SAM3 model finds and masks those elements across your frames. This is the fastest way to remove the operator and obvious moving objects.

Keyword masking · SAM3Type what you want gone
PromptThreshold50%
1 element masked across 240 frames

Object picker (SAM3)

By default every match for your keyword is masked. Turn on Object picker in the SAM3 options when you only want some of them — one person out of a crowd, or every car except the one parked in the scene.

  1. Enter your keywords and press Detect this frame. Each match gets a box.
  2. Click the boxes you want. Click again to deselect. Where boxes overlap, the click picks the smallest one, so you can grab a person standing in front of a larger match.
  3. Press Run Mask. SAM tracks your picks through the whole clip.
Mode What gets masked
Mask selected (default) Only the boxes you picked (shown green).
Exclude selected Every keyword match except your picks (shown red) — they stay in the scene.
Object picker · SAM3Keyword “person” — click the boxes to pick

Mask selected — only your picks (green) are masked; the other matches stay.

Picked
1 of 4 matches masked · tracked through the clip

Frames are about a second apart, so the track can lose an object. If that happens, scrub to where it was lost and pick it again there — every frame you pick on becomes a new anchor. In a batch, pick on each video too; a video with no picks gets every match masked (Exclude) or nothing masked (Mask selected).

Switching the picker on or off, or changing the mode, clears the SAM mask, so detect again afterwards. The first run loads the tracking model and takes 10–30 seconds longer to start.

Paintbrush

Paint masked areas by hand for anything the keyword model misses, or for one-off objects that don't have a clean label. Useful for reflections, tripods and stray clutter.

Circle masks

Mask out the area outside the fisheye image circle — the black corners and lens barrel that dual-fisheye captures include. They come with per-camera presets and can be nudged to fit your exact lens.

Exporting masks

You can export just the masks if you want to reuse them in another tool or keep a record of what was removed from each frame.

Long clips and resuming

Keyword masks are saved as they're made. If a masking run stops — you press stop, the computer restarts, or it runs out of memory — press Run Mask again and it continues where it stopped. The log says Reusing N cam-frame masks saved by the previous run.

Saved masks are thrown away when you change the keyword, threshold, mode or frames, and Clear All Masks deletes them. Object-picker runs and Detect this frame always start from the beginning.

If a very long clip or batch runs out of memory, the step stops with Out of memory while masking. Extract fewer frames (a lower frame rate) and run the Mask step again.