A newly posted paper on Hugging Face introduces RGBD20K, a large-scale benchmark designed to advance RGB-D semantic segmentation. According to the authors, the dataset is built to support more robust and general segmentation models by combining a broad category range with high-quality annotations.
The benchmark is described as having three main properties. First, an expanded semantic space: it covers 160 fine-grained categories, which the authors say far exceeds the category diversity of widely used RGB-D benchmarks such as NYUv2 (40 classes) and SUN RGB-D (37 classes). Second, a larger scale: RGBD20K provides 20,000 RGB-D image pairs, offering a substantially bigger training resource than existing benchmarks. Third, high-fidelity annotation: the team re-evaluated and corrected existing labels to address long-standing annotation noise, producing a cleaner ground-truth foundation.
Beyond the dataset, the paper proposes a novel Score-Purified Fusion (SPF) method. The authors state that SPF achieves state-of-the-art performance across all evaluated benchmarks, which they present as evidence that high-quality multimodal information can be leveraged effectively for RGB-D semantic segmentation.
The dataset and pretrained models are said to be publicly available, with the dataset hosted at a GitHub repository linked in the paper. The Hugging Face page also includes automated recommendations of related papers from the Semantic Scholar API, covering topics such as efficient RGB-D segmentation, RGB-D video salient object detection, and multimodal 3D object detection.