Official PyTorch implementation for "NeRFCom: Feature Transform Coding Meets Neural Radiance Field for Free-View 3D Scene Semantic Transmission".
As described in the paper, NeRFCom training is divided into three stages:
-
Pretrain the NeRF module: You can either train from scratch or use publicly available pretrained models. We recommend using models with geometric priors, such as DVGO, TensoRF, or K-Planes.
-
Train the Encoder and Decoder modules: We recommend performing training in two sub-stages (w/o and w/ channel) to stabilize convergence and improve performance.
-
Joint training with all components: After the above two stages, jointly fine-tune the entire NeRFCom pipeline for best performance.
Maybe Help:
- The training mode for each stage currently needs to be manually adjusted in the code.
- You are encouraged to monitor the learning rate schedules and gradient flows to verify that only the intended modules are updated during each phase.
- This repository is currently undergoing reorganization. Partial code has been uploaded, and more updates will be released soon.
This project builds upon the excellent work of the following repositories:
We sincerely thank the authors of these projects for their contributions.