SeeThrough3D: 10K-100K Multi-Object Scenes with Occlusion-Aware 3D Layouts

Comprising between 10,000 and 100,000 multi-object scenes designed for training text-to-image models with explicit occlusion modeling. Developed by va1bhavagrawa1 for a CVPR 2026 publication, it introduces an Occlusion-aware 3D Scene Representation (OSCR) using translucent 3D boxes.

Use Cases

• Training text-to-image models using OSCR (Occlusion-aware 3D Scene Representation) for spatial control
• Benchmarking occlusion handling in 3D layout-conditioned generation tasks
• Fine-tuning controllable generation models using the translucent 3D box annotations

Strengths

• Scale of 10,000 to 100,000 records
• Specialized OSCR labels for occlusion modeling
• Peer-reviewed research foundation (CVPR 2026)

Limitations

• Schema details for the JSON layout files are not explicitly listed
• Narrow focus on 3D layout-conditioned generation

Provenance

Source

CVPR 2026 Paper: SeeThrough3D: Occlusion Aware 3D-Control in Text-to-Image Generation

Freshness

Last updated March 2026.