wenjingbian/structure_data

该数据集名为structure_data — unified-schema 3D reconstruction datasets，主要用于图像到3D结构预测的任务，是Qwen3-VL-4B-Instruct的SFT数据。数据集包含三个子集：v11canon（单物体，如椅子、桌子、柜子等）、scene_v3（多物体场景）和physx_v11canon（PartNet-Mobility可动部件物体，旋转至v11手性）。每个子集都有训练和测试数据，以及对应的图像文件。数据格式为JSONL，每行包含图像路径和对话内容，对话中人类提问和AI回答的结构化3D信息。3D信息的GT JSON包含材料（材料家族和描述）、节点（父节点、材料、质心、范围等）和关系（对称、实例等）等字段。图像约定包括：第一张为带材质的渲染图，第二张为分割图，其余为额外视角图。坐标系约定为Y轴向上，X轴向前，单位立方体归一化。

The dataset is named structure_data — unified-schema 3D reconstruction datasets and is primarily used for image-to-3D structure prediction tasks, serving as SFT data for Qwen3-VL-4B-Instruct. It consists of three subsets: v11canon (single objects like chairs, tables, cabinets, etc.), scene_v3 (multi-object scenes), and physx_v11canon (PartNet-Mobility articulated objects, rotated to v11 chirality). Each subset includes training and test data along with corresponding image files. The data format is JSONL, with each row containing image paths and conversational content, where humans ask questions and AI responds with structured 3D information. The GT JSON for 3D information includes fields such as materials (family and description), nodes (parent, material, centroid, extents, etc.), and relations (symmetric, instance, etc.). Image conventions specify: the first image is a rendered view with full materials, the second is a segmentation map, and the rest are additional views. Coordinate conventions are Y-up, +X-front, normalized to a unit cube.