SceneTrans: Paired Synthetic Indoor Scene Images for Change Detection

A synthetic dataset of paired indoor scene images for scene change detection research. Each pair shows the same scene before and after an object-level edit, rendered with NVIDIA Isaac Sim 5.1.0. The dataset includes RGB, depth, and segmentation modalities plus natural-language descriptions and corresponding action labels.

Use Cases

• Training scene change detection models based on paired before/after images.
• Developing models for natural-language description of scene edits based on the provided text descriptions.
• Benchmarking multimodal AI systems using the combined RGB, depth, and segmentation data.
• Researching synthetic data generation and its application in computer vision tasks.

Strengths

• Provides paired image sets (A/B pairs) specifically designed for change detection tasks.
• Includes multiple data modalities per scene: RGB, depth, and segmentation.
• Each image pair is accompanied by a natural-language description and an action label.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

ShenGroup

Collection Method

Synthetically generated using NVIDIA Isaac Sim 5.1.0 from USD assets.

Freshness

Last updated 2026-05-29 20:05:07; freshness should be verified.