Haian Jin (@Haian_Jin)
2025-12-12 | โค๏ธ 107 | ๐ 10
Impressive results! The 3D field has exhausted almost all publicly available 3D datasets (~30 for now). Self-supervised E-RayZer provides a vital new path for further scaling data-driven 3D models. The next milestone will likely be how to include dynamic videos into the pipeline.
์ธ์ฉ ํธ์
Hanwen Jiang (@hanwenjiang1)
(1/N) Will this be the BERT/GPT moment for 3D vision๏ผ Finally, unsupervised pre-training for 3D works.
Led by @qitao_zhao , we present E-RayZer โ a fully self-supervised 3D reconstruction model that: ๐ฅMatches or surpasses supervised methods like VGGT ๐Learns transferable 3D representations, outperforming CroCo, VideoMAE, and DINO ๐Scales with more unlabeled data
A new recipe for scalable 3D foundation models.
๐ฌ ์์
