OpenShape: Scaling Up 3D Shape Representation Towards Open-World Understanding

1University of California San Diego 2Shanghai Jiao Tong University 3Qualcomm
teaser image.

Left: Zero-shot 3D shape classification on the Objaverse-LVIS (1,156 categories) and ModelNet40 datasets (40 common categories). Right: Our shape representations encode a broad range of semantic and visual concepts. We input two 3D shapes and use their shape embeddings to retrieve the top three shapes whose embeddings are simultaneously closest to both inputs.


We introduce OpenShape, a method for learning multi-modal joint representations of text, image, and point clouds. We adopt the commonly used multi-modal contrastive learning framework for representation alignment, but with a specific focus on scaling up 3D representations to enable open-world 3D shape understanding. To achieve this, we scale up training data by ensembling multiple 3D datasets and propose several strategies to automatically filter and enrich noisy text descriptions. We also explore and compare strategies for scaling 3D backbone networks and introduce a novel hard negative mining module for more efficient training. We evaluate OpenShape on zero-shot 3D classification benchmarks and demonstrate its superior capabilities for open-world recognition. Specifically, OpenShape achieves a zero-shot accuracy of 46.8% on the 1,156-category Objaverse-LVIS benchmark, compared to less than 10% for existing methods. OpenShape also achieves an accuracy of 85.3% on ModelNet40, outperforming previous zero-shot baseline methods by 20% and performing on par with some fully-supervised methods. Furthermore, we show that our learned embeddings encode a wide range of visual and semantic concepts (e.g., subcategories, color, shape, style) and facilitate fine-grained text-3D and image-3D interactions. Due to their alignment with CLIP embeddings, our learned shape representations can also be integrated with off-the-shelf CLIP-based models for various applications, such as point cloud captioning and point cloud-conditioned image generation.

Zero-Shot 3D Shape Classification

comparison image.

Multi-Modal 3D Shape Retrieval

3D shape retrieval from image (left, mid) and point cloud (right).
Text-input 3D shape retrieval. In each row, we show input texts on the left and two retrieved shapes for each text on the right. OpenShape embedding encodes a wide range of visual and semantic concepts and enables (a) retrieval of fine-grained subcategories (first two rows), and (b) control of attributes (e.g., color, shape, style) and their combinations (last two rows).

Shape-Conditioned Multimodal Generation

(a) Point cloud captioning. (b) Point cloud-conditioned image generation. Our learned 3D shape embeddings can be integrated with off-the-shelf pretrained CLIP-based models (e.g., captioning and image generation models) to support various cross-modal applications.

More Examples

Image-input 3D shape retrieval. Each triplet shows input image and two retrieved 3D shapes.
Point cloud-input 3D shape retrieval. Each triplet shows input point cloud and two retrieved 3D shapes.
Point cloud captioning.
Point cloud conditioned image generation. Each pair shows the input point cloud and the generated image.


      title={OpenShape: Scaling Up 3D Shape Representation Towards Open-World Understanding}, 
      author={Minghua Liu and Ruoxi Shi and Kaiming Kuang and Yinhao Zhu and Xuanlin Li and Shizhong Han and Hong Cai and Fatih Porikli and Hao Su},