MGMA: Mesh Graph Masked Autoencoders for Self-supervised Learning on 3D Shape

Zhangsihao Yang

Keywords: mesh graph, self-supvervised learning, masked autoencoder, attention

TL;DR: We introduce a self-supervised learning model to extract face nodes and global graph embeddings on meshes.

Abstract: We introduce a self-supervised learning model to extract face nodes and global graph embeddings on meshes. We define a graph masking on a mesh graph composed of faces. We evaluate our model on shape classification and segmentation benchmarks. The results suggest that our model outperforms prior state-of-the-art mesh encoders: In ModelNet40 classification task, it achieves an accuracy of 89.8% and in ShapeNet segmentation task, it achieves a mean Intersection-over-Union (mIoU) of 78.5. Further, we explore and explain the correlation between test and training masking ratios on MGMA. And we find best performances are obtained when mesh graph masked autoencoders are trained and evaluated under different masking ratios. Our work may open up new opportunities to address label scarcity and improve the learning power in geometric deep learning research.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning