joshir/3D-Scene-Segmentation-HQ

3D Segmentation HQ Dataset

The 3D Segmentation HQ dataset is a curated collection of 5 real-world scenes with high-quality object segmentation masks designed for research in 3D scene understanding, editing, and rendering. This dataset improves upon existing benchmarks by providing cleaner and more consistent object masks across multiple views, enabling reliable evaluation and training for tasks such as:

• 3D semantic segmentation
• Object-level scene editing (e.g., removal, recolorization)
• 3D Gaussian Splatting with semantic supervision

---

Dataset Composition Total Scenes: 5 real-world scenes

Source Datasets:

3 scenes are adapted from the LERF dataset (https://arxiv.org/abs/2303.09553)

2 scenes are adapted from the LLFF dataset (https://huggingface.co/datasets/nerfbaselines/nerfbaselines-data)

Each scene is re-processed with HQ object segmentation masks to ensure multi-view consistency, reducing the noise and inconsistencies commonly found in previous datasets.