sensenova/SenseNova-Vision-Corpus-50M
收藏资源简介：
SenseNova Vision Corpus 50M（SN-VC-50M）是一个大规模多模态视觉语料库，专为统一训练多种视觉理解和几何导向任务而设计。该数据集旨在解决现有公共视觉数据集的常见局限性，例如注释不完整、跨任务不一致或无法直接解码为训练就绪的多模态监督。SN-VC-50M将开源视觉数据组织为四个任务家族：结构化视觉理解、分割、密集几何预测和多视图视觉几何。在这些家族中，该版本包含73个数据集-任务条目，覆盖10种任务类型。具体规模包括：结构化视觉理解18.9M帧、分割1.3M帧、密集几何预测17.3M帧、多视图视觉几何12.5M帧。为构建训练兼容的监督信息，数据集使用任务专门化的处理流程：对于结构化理解，采用Rex-Omni数据构建流程生成检测和OCR风格的样本；对于密集几何，利用MoGe-2来稠密化稀疏深度和表面法线注释并提高场景多样性；对于多视图场景，使用LingBot-Depth补充不完整的稀疏深度信息；对于分割，应用严格的对齐检查以确保文本区域描述、颜色图例和分割掩码之间的一致性。为避免重新分发公共源数据集中重复的原始RGB图像，JSONL训练示例保留相应的相对文件路径，而不是复制所有原始RGB资源。用户需要将本地数据集根目录与JSONL文件中记录的图像文件路径对齐，以正确加载视觉输入。
SenseNova Vision Corpus 50M (SN-VC-50M) is a large-scale multimodal visual corpus designed specifically for unified training of diverse visual understanding and geometry-oriented tasks. This dataset aims to address common limitations of existing public visual datasets, such as incomplete annotations, cross-task inconsistencies, and inability to be directly decoded into training-ready multimodal supervision. SN-VC-50M organizes open-source visual data into four task families: structured visual understanding, segmentation, dense geometry prediction, and multi-view visual geometry. Within these families, this version includes 73 dataset-task entries covering 10 task types. The specific scale is as follows: 18.9M frames for structured visual understanding, 1.3M frames for segmentation, 17.3M frames for dense geometry prediction, and 12.5M frames for multi-view visual geometry. To create training-compatible supervision information, the dataset adopts task-specialized processing workflows: For structured visual understanding, the Rex-Omni data construction pipeline is used to generate detection and OCR-style samples; For dense geometry prediction, MoGe-2 is utilized to densify sparse depth and surface normal annotations and improve scene diversity; For multi-view visual geometry scenarios, LingBot-Depth is employed to supplement incomplete sparse depth information; For segmentation, strict alignment checks are applied to ensure consistency among text region descriptions, color legends, and segmentation masks. To avoid redistributing duplicate raw RGB images from public source datasets, JSONL training examples retain their corresponding relative file paths instead of copying all raw RGB resources. Users must align their local dataset root directory with the image file paths recorded in the JSONL files to properly load visual inputs.
