implementation details for propagation modules in ViTDet

[wenmengzhou]

We have implemented ViT-Det with global propagation in #3 , but not provided the implementation of convolutional propagation, which involves MAE pretraining. As we have implemented MAE, it is straghtforward to try to implement this module before the official code is release. @tuofeilunhifi

The mainly description is as follows

[wenmengzhou]

@tuofeilunhifi "Our backbone adaptation is simple and makes detection fine-tuning compatible with global self-attention pre-training. As stated, it is not necessary to
redesign the pre-training architectures." from this statement, we may only add conv propagation during the finetuning stage, and leaving MAE preatraining as it is ？

[Michaell109]

Yeah, my understanding of the statement is the same as yours. The conv propagation is only needed in finetuning stage. By the way, would you like to share the mAP of ViTDet (base) you implemented? It would be very helpful for me. Thanks very much.

[wenmengzhou]

sure, as we finished refactoring some code, we will release both code and models, maybe expected at next week

[Michaell109]

Thanks! I am looking forward to it.

[tuofeilunhifi]

Yeah, my understanding of the statement is the same as yours. The conv propagation is only needed in finetuning stage. By the way, would you like to share the mAP of ViTDet (base) you implemented? It would be very helpful for me. Thanks very much.

The accuracy of conv aggreation we reproduced is 42.13, so we guess that pre-training of conv aggregation is required