D-RE10K: 147,422 Frames for Dynamic Real-Estate View Synthesis

D-RE10K contains 15,467 training clips and 147,422 frames of real-estate walkthrough videos featuring dynamic objects. Developed by uva-cv-lab and released in 2026, the collection is designed for training and evaluating self-supervised large view synthesis models in non-static environments.

Use Cases

• Training novel view synthesis models using the 147,422 training frames
• Evaluating dynamic scene reconstruction accuracy using the 1,541 mask test frames
• Self-supervised 3D reconstruction research using real-estate walkthrough video sequences

Strengths

• 15,467 training clips containing 147,422 frames
• Includes a dedicated mask test split of 76 clips for dynamic scene evaluation
• Curated specifically for dynamic object handling in real-estate environments

Limitations

• Gated access requires user approval from the authors
• Limited to real-estate walkthrough perspectives which may not generalize to all outdoor dynamic scenes

Provenance

Source

uva-cv-lab, associated with Arxiv publication 260110716

Collection Method

Curated from real-estate walkthrough videos

Freshness

Last updated March 2026.