search

ActiveVisionLab / DFNet

DFNet: Enhance Absolute Pose Regression with Direct Feature Matching (ECCV 2022)
https://dfnet.active.vision
MIT License
95 stars 9 forks source link

Question about the model training process #9

Closed LZL-CS closed 11 months ago

LZL-CS commented 1 year ago

Hi @chenusc11, in the pipeline of your DFNet paper, it seems that you use estimated pose P(hat) for Histogram-assisted NeRFH model to generate a rendered image, which is similar to your previous paper Direct-PosNet training process.

image

image

However, in the DFNet code, I didn't find any place where you use the estimated pose for Histogram-assisted NeRFH. While in the Direct-PosNet code, you first train a PoseNet for camera pose estimation, then use this pre-trained model for NeRF model. May I know the reason for the difference in training between DFNet and Direct-PosNet? Maybe I miss something important, hope for your reply.

Kind regards, Thanks!

chenusc11 commented 11 months ago

Not sure I understand your question. However, DFNet itself could estimate more accurate camera poses than PoseNet. I'm closing the issue for now. If you want to further elaborate on this issue, please feel free to do so.