SceneSplat-7K was released in 2025 by the University of Amsterdam, the Computer Vision Laboratory of the Swiss Federal Institute of Technology in Zurich, INSAIT of Sofia University, and other institutions. It is currently the largest and highest-quality indoor scene 3D Gaussian Splats (3DGS) dataset. The related paper results are "SceneSplat: Gaussian Splatting-based Scene Understanding with Vision-Language Pretraining", which aims to promote the understanding and semantic reasoning capabilities of vision-language pre-training models in real indoor 3D scenes.

This dataset compiles 3DGS data generated from multiple data sources, totaling approximately 9,000 original scenes. Of these, 7,916 scenes were processed into 3DGS, each containing an average of 1.42 million Gaussian points, for a total of 11.27 billion Gaussian points. Based on these, 4,114 high-quality scenes were selected for pre-training. The dataset also includes 4.72 million RGB frames, demonstrating excellent reconstruction quality: average PSNR = 29.64 dB, average depth error = 0.035m, average SSIM = 0.897, and average LPIPS = 0.212.