SwitchSplat Dataset

Twelve real scenes, each filmed casually while someone changes the state of objects in it, such as opening doors, drawers, cupboards and fridges, or the doors and hood of a car. The camera was a consumer Insta360 X4 dual-fisheye 360° camera. This is the benchmark of SwitchSplat: Learning Discrete Object States from Casually Captured Scenes (SIGGRAPH Asia 2026).

Paper · Project page · Interactive viewer

Front-lens frames with an object in its changed state. Top row: Balcony, Car, Kitchen 1, Flat 1. Middle: Flat 2, Flat 3, Chair, M Kitchen 1. Bottom: M Kitchen 2, M Kitchen 3, Kitchen 2, Oslo.

Scenes

The standard set has 10 scenes (8 indoor, 2 outdoor) with 3–9 interactions each; the three M Kitchen scenes have overlapping interactions. The golden set (Kitchen 2, Oslo) adds controlled recaptures for testing new camera paths and combinations of object states never seen together during the capture.

Scene	Folder	Set	Setting	Interactions	Images	Test images	Pinhole zip	Fisheye zip
Balcony	balcony	standard	outdoor	5	1,858	200	5.6 GB	5.8 GB
Car	car	standard	outdoor	4	2,310	262	6.7 GB	6.7 GB
Kitchen 1	kitchen1	standard	indoor	9	4,252	476	6.6 GB	7.8 GB
Flat 1	flat1	standard	indoor	5	2,662	292	4.6 GB	5.2 GB
Flat 2	flat2	standard	indoor	9	4,614	528	8.4 GB	9.4 GB
Flat 3	flat3	standard	indoor	9	4,230	486	6.9 GB	8.0 GB
Chair	chair	standard	indoor	3	1,478	172	1.6 GB	2.1 GB
M Kitchen 1	m_kitchen1	standard	indoor, concurrent	3	1,056	115	1.5 GB	1.8 GB
M Kitchen 2	m_kitchen2	standard	indoor, concurrent	3	1,372	150	2.0 GB	2.3 GB
M Kitchen 3	m_kitchen3	standard	indoor, concurrent	3	1,648	184	2.5 GB	3.0 GB
Kitchen 2	kitchen2	golden	indoor	3	2,332	198 + 598	3.4 GB	4.0 GB
Oslo	oslo	golden	indoor	3	1,964	160 + 600	2.8 GB	3.3 GB

Images count both lenses. Test images are the standard test frames; for the golden set, the second number is the recaptured novel-camera + novel-combination frames. scenes.csv has the full counts.

Download

Each scene is one zip per version. Archives unpack into <version>/<scene>/.

pip install -U huggingface_hub
# one scene
hf download xieyizheng/switchsplat pinhole/kitchen2.zip SHA256SUMS --repo-type dataset --local-dir switchsplat
# a whole version (pinhole: 53 GB, fisheye: 59 GB)
hf download xieyizheng/switchsplat --repo-type dataset --include "pinhole/*" --include SHA256SUMS --local-dir switchsplat

cd switchsplat
sha256sum -c --ignore-missing SHA256SUMS
for f in pinhole/*.zip; do unzip -q "$f"; done

Two versions

• pinhole/: the version used for the paper's main experiments. Frames are undistorted by COLMAP into pinhole images (PINHOLE cameras).
• fisheye/: the same frames before undistortion (OPENCV_FISHEYE cameras). Camera poses, frame names, annotations and splits are identical to the pinhole version, and the colours were matched to it. The black area outside the image circle is not part of the image.

Folder layout

<scene>/
  images_2/camera1/frame_0001.png                 front lens, 1440x1440 RGB
  images_2/camera2/frame_0001.png                 back lens
  masks_2/masks-20_2/camera1/frame_0001_mask.png  camera-holder mask
  sparse/0/                                       COLMAP model
  test_frames.txt
  interaction_times.txt
  transition_frames.txt
  vlm_prediction.txt

• Images. Frames were taken from the video at 4 fps; frame_XXXX is the frame number, shared by the two lenses. Images are at half resolution, while the COLMAP model is at the original 2880x2880, so halve the intrinsics (downsample factor 2, as the images_2 name says). A few frames per lens could not be registered by COLMAP and are left out.
• Camera-holder masks. Made with the FullCircle masking pipeline. Pixels above 127 are the person holding the camera; SwitchSplat ignores them in training and in its metrics.
• Moving-background masks. Pinhole Car and Balcony also have masks_2/masks-skytreeplant-20_2/ (sky, trees and plants), used in the supplement to score the outdoor scenes without them.
• sparse/0/ holds cameras.bin, images.bin and points3D.bin. Where COLMAP wrote them, models also have rigs.bin and frames.bin (its rig format): pinhole Chair, M Kitchen 1, M Kitchen 2 and M Kitchen 3; all fisheye scenes except Kitchen 1, Flat 3 and Oslo.

Annotations and splits

• interaction_times.txt: one line per interaction, in time order: start end frame numbers of the period in which the object is in its changed state (for example, the door is open).
• transition_frames.txt: one line per interaction: on_start on_end off_start off_end, the frames in which the object is moving into and out of its changed state.
• vlm_prediction.txt: the interaction periods predicted automatically by GPT-5.6 Sol from the front-lens frames (start end per line). SwitchSplat uses this file, not the annotation, to set up its object slots.
• test_frames.txt: the held-out images, as paths under images_2/. Everything else is training data.

The evaluation splits of the paper:

• Standard test: every 8th frame of the capture, both lenses, except frames during a transition.
• Base-state test: standard test frames outside every interaction period, so all objects are in their starting state.
• Novel camera (golden set): frame_test_novel_view_<state>_<n>.png, recaptured along new camera paths with object states that do occur in the capture.
• Novel state combination (golden set): frame_test_novel_combo_<state>_<n>.png, combinations of object states that never occur together in the capture.

<state> has one digit per interaction, in the order of interaction_times.txt; 1 means that object is in its changed state. The recaptured frames are registered in the same COLMAP model and are never used for training.

Limitations

• The camera holder is masked, not removed or blurred, so people are visible in the images, mostly in the back lens. The masks are automatic and occasionally miss part of a hand or arm.
• The outdoor scenes also change for reasons unrelated to the objects: wind in the trees, moving clouds and distant traffic.

License

CC BY 4.0.

Citation

@inproceedings{xie2026switchsplat,
  title     = {SwitchSplat: Learning Discrete Object States from Casually Captured Scenes},
  author    = {Xie, Yizheng and Foroutan, Yalda and Shramko, George and Rebain, Daniel
               and Govindarajan, Shrisudhan and Tagliasacchi, Andrea},
  booktitle = {SIGGRAPH Asia 2026 Conference Papers},
  year      = {2026}
}

We thank Zhihan Zhang, Egor Fokin, and Mingrui Zhao for contributing scenes and objects used in the captures.