NVlabs/Deep_Object_Pose

(Question) About weight & fine tuning

Opened this issue · 3 comments

umegan commented

Hi, after a week of generating data and training the model, I obtained a pretty good result on 6D pose estimation. However, it takes quite a long time, so I want to test fine-tuning on DOPE and have a question about the weights and fine-tuning.

・Do you have a list of object image paths to the weight

・Does fine-tuning on objects with similar shapes yield good results?

TontonTremblay commented

This is amazing results. I am so sorry for the amount of time / work you had to put into this. I am afraid that I never spent anytime trying to build a model that could easily be fine-tuned.

umegan commented

Oh, I see. What about transfer learning on the pretrained DOPE model?

  1. Download the weights and save them to the weights folder, i.e., ~/catkin_ws/src/dope/weights/.

Does it learn faster on similar shapes compared to training from scratch?

TontonTremblay commented

I am not sure, I never tried it, let me know if you end up trying.

  • 👀2