SceneTrans

SceneTrans is a synthetic dataset of paired indoor scene images (A/B pairs) for scene change detection / "spot-the-difference" research. Each pair shows the same scene before and after an object-level edit — an object being added, moved, removed, or left unchanged — together with a natural-language description and the corresponding action.

All scenes are rendered with NVIDIA Isaac Sim 5.1.0 from USD assets, and come with RGB, depth, and segmentation modalities plus ground-truth metadata about what changed.

💻 Generation scripts, annotation notebooks, and full documentation are on GitHub: ShenGroup/SceneTrans.

Example

A snapshot of representative scenes and A/B pairs in the dataset:

Scenes

The data is distributed as per-scene .zip archives (the number in each filename is the number of A/B pairs), spanning 8 indoor scenes and roughly 14,605 pairs in total.

Archive	Scene	A/B pairs
kitchen_sim_2300.zip	Kitchen	2,300
studyroom_desk_sim_2400.zip	Study room – desk	2,400
bedroom1_bed_sim_2390.zip	Bedroom – bed	2,390
livingroom_sim_1990.zip	Living room	1,990
livingroom_shelf_sim_1925.zip	Living room – shelf	1,925
childrenroom_desk_sim_1660.zip	Children room – desk	1,660
diningroom_sim_1140.zip	Dining room	1,140
bedroom1_shelf_sim_800.zip	Bedroom – shelf	800
assets.zip	USD scene/object assets (source environments)	–

Change Types

Every pair belongs to one of four object-level transformations:

• Add — an object absent in frame A is present in frame B.
• Move — an object is relocated (and possibly re-oriented) between the two frames.
• Remove — an object present in frame A is absent in frame B.
• Unchanged — both frames come from the same rendered state (a valid no-change sample).

License

Released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).