Tag
The paper identifies a theoretical incompatibility between 3D point cloud masked autoencoding and SE(3)-equivariant networks, proving that standard masking triggers a topological collapse, and introduces SIEVE, an equivariant self-supervised framework using a spherical projection mask to preserve global pose while obfuscating local semantics, improving downstream generalization to unseen shape categories.
提出一种两遍式动态 k-NN 图调度方法，通过阻尼不动点迭代与信任域补偿，在 SE(3) 等变结构预测中以 1.5×/1.76× 加速而不损失精度，并给出非翻转证书与近间距解析定律。