只想重新训练第三阶段（论文中的C阶段）

[Wenhua Liao]

您好，非常高兴您能分享您具有启发性和富有成果的工作，目前我想将我的部分工作迁移到您的方法上来，但是我只需要重新训练第三阶段，请问是否可以分享相关的资源使得第三阶段可以独立训练，包括并不限于第二阶段训练结束的模型参数，viewpoint codebook等等，十分感谢，祝工作顺利，论文多多！！

[Dingding]

Hi, thanks for your interest in our work!

Sorry, I don't completely understand your question. Could you provide more details?

The entire pipeline was trained end-to-end instead of stage-by-stage, so there is no separate pre-trained model weight. But you can extract the weight parameters of each separate stage (e.g., the viewpoint encoder in stage B, the regression module in stage C, or the verification module in stage D) from the pre-trained model (https://drive.google.com/drive/folders/16f2xOjQszVY4aC-oVboAD-Z40Aajoc1s).

The third stage (i.e., stage C ) is a lightweight network consisting of a single Conv2D + two FC layers. Given a pair of feature maps as input, This network is trained to regress the relative in-plane rotation of the input pair. We don't need the viewpoint codebook for training as the training data is synthesized on the fly.