"MGPC-1M"

"MGPC-1M is the official dataset for our paper \"MGPC: Multimodal Network for Generalizable Point Cloud Completion With Modality Dropout and Progressive Decoding\". MGPC-1M is a large-scale multimodal benchmark for generalizable point cloud completion. It contains over 1,000 object categories and more than one million paired samples constructed via an automated pipeline. For each sample, we provide (1) a partial point cloud simulated from rendered depth by back-projection with sensor-like noise, (2) the corresponding RGB observation rendered from virtual cameras distributed across multiple viewpoints, (3) a text description automatically derived from object tags or category labels, and (4) a complete reference point cloud uniformly sampled from the underlying mesh. To better match real deployment where ground-truth geometry is unavailable, MGPC-1M adopts input-centric normalization and stores the associated normalization parameters for each pair. The training split is built from large-scale synthetic Objaverse and ShapeNet, while the test split comprises 1,031 real-world instances from Google Scanned Objects and is entirely disjoint from the training data, enabling rigorous cross-domain generalization evaluation. To ensure data quality at scale, a VLM-assisted filtering stage removes low-quality meshes (e.g., corrupted files or unrealistic geometry), complemented by manual spot checks. MGPC-1M supports research on single- and cross-modal completion, multimodal fusion, robustness under realistic sensing conditions, and large-scale benchmarking for 3D reconstruction and graphics\/robotics applications. For more details, please refer to our paper and the official code https:\/\/github.com\/L-J-Yuan\/MGPC."