[CNN architecture support]

[CaptainEven]

Out of curiosity, may i ask is there any possibility to make a version of CNN based CROCO self-supervised pipeline?

[xjcvip007]

Maybe 《MC-JEPA: A Joint-Embedding Predictive Architecture for Self-Supervised Learning of Motion and Content Features》can meet your requirements.

[PhilippeWeinzaepfel]

Hi,

Masked Image Modeling (MIM) methods in general are well desiged for patch-based architectures such as ViTs.
There have been some attemps to extend MIM to CNNs, eg Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN or ConvMAE: Masked Convolution Meets Masked Autoencoders.
Such approaches could most likely be successfully integrated into CroCo but we are planning to work on that in the future.

Best
Philippe

[CaptainEven]

Thansk for the reply!