SeokjuLee / Insta-DM

Learning Monocular Depth in Dynamic Scenes via Instance-Aware Projection Consistency (AAAI 2021)
Other
219 stars 27 forks source link

pre-training weights #5

Open chaoxingxi opened 3 years ago

chaoxingxi commented 3 years ago

How are the pre-training weights of each sub-network obtained? How are the sub-networks supervised learning?

SeokjuLee commented 3 years ago

The whole joint-training pipeline is optimized by self-supervision, as defined in Eq. (1) in our paper. Could you please clarify your question?

chaoxingxi commented 3 years ago

As stated in your paper, use data with segmentation for unsupervised training. At the beginning of training, do the weights of the DispResNet, EgoPoseNet and the ObjPoseNet use randomly initialized initial values? Or use supervised learning to train the weights, and then use the unsupervised framework to fine-tune? Thank you ~