Recently, Associate Professor Zhenxin Zhang, a young teacher from the School of Resources and Environment, published a paper titled "Cross-Domain Point Cloud Completion for Multi-Class Indoor Incomplete Objects Via Class-Conditional Gan Inversion" in ISPRS Journal of Photogrammetry and Remote Sensing, an official journal of the International Society for Photogrammetry and Remote Sensing (ISPRS) and a top international journal in the field of photogrammetry and remote sensing, which breaks through the technical bottleneck of multi-class indoor 3D object completion. Associate Professor Zhenxin Zhang is the first author of the paper, and Capital Normal University is the sole corresponding unit. The co-authors include Leng Siyi, a master's student of Capital Normal University in 2020, and Professor Liqiang Zhang from Beijing Normal University.

The three-dimensional point cloud obtained from the real indoor scene scanning is often incomplete, sparse and noisy due to the limited viewing angle and occlusion between objects (as shown in Fig. 1). The key to point cloud completion is not only to ensure the data integrity, but also to promote high-quality implementation of tasks such as point cloud classification, segmentation, object detection, and 3D reconstruction.

Zhenxin Zhang, a young teacher from the School of Resources and Environment, led a research team to break the constraints of previous point cloud completion methods, no longer focusing on synthetic data, let alone the need to model each object class separately, breaking through the limitations of completing a single class of three-dimensional objects. The team creatively proposed an efficient class-conditional GAN inversion framework (as shown in Fig. 2 and Fig. 3). By training class-conditional GAN in the synthetic data domain, learn the shape prior of multi-class objects, find the optimal latent code for each incomplete object point cloud in the GAN latent space, and then generate a complete shape using this latent code to best recover the missing parts in the given real data domain. This method overcomes the limitations of obtaining paired incomplete and complete point cloud data in real scenes, and realizes the high-quality completion of multi-class objects in cross domain real indoor 3D scenes.

This method has advanced the level of cutting-edge methods and broken through the technical bottleneck of achieving high-quality completion on multi-class indoor incomplete objects (as shown in Fig. 4 and Fig. 5). It also provides a technical foundation for the construction of the digital economy and the real 3D China, further enriching and expanding the information completeness theory of earth science, information science, and interdisciplinary fields. It has been supported by the National Natural Science Foundation of China and other projects.

ISPRS Journal of Photogrammetry and Remote Sensing is a top tier journal in the Chinese Academy of Sciences, with an Instantaneous Impact Factor of 12.7. In recent years, the School of Resources and Environment has encouraged and guided young teachers to conduct scientific research and cultivate high-quality talents in response to major national needs. The relevant research achievements have been applied in national major projects such as Mars exploration and the construction of a real 3D China, achieving certain social benefits and contributing to construction of "Double First-Class" of our university and the implementation of "Climbing Program".

Paper information:

Zhenxin Zhang, Siyi Leng, Liqiang Zhang, Cross-domain point cloud completion for multi-class indoor incomplete objects via class-conditional GAN inversion, ISPRS Journal of Photogrammetry and Remote Sensing, Volume 206, 2023, Pages 118-131.

Paper website:

https://www.sciencedirect.com/science/article/pii/S0924271623003064.