search

sconlyshootery / FeatDepth

This is the offical codes for the methods described in the "Feature-metric Loss for Self-supervised Learning of Depth and Egomotion".
MIT License
247 stars 28 forks source link

online refinement #66

Closed fxl-fxl closed 3 years ago

fxl-fxl commented 3 years ago

Thank you for your excellent work. I have a question. Does this online refinement use the 697 test sets to train for 20 epochs and then test the accuracy on the 697 test sets? But in other work(https://arxiv.org/abs/2104.14540), they performed 20 iterations on a single sample of the test set one by one, and loaded the same trained model for the training each sample. Do you have any suggestions

sconlyshootery commented 3 years ago

Our online refinement keeps the same settings with other works, just like you described.

brandleyzhou commented 3 years ago

I think you are right, they (https://arxiv.org/abs/2104.14540) finetune models for every sample in a test set, which means one model for one sample. Finally, they will have 697 models to evaluate. But in this work, after TTR, there is a single model. There are two different TTR techniques.

Thank you for your excellent work. I have a question. Does this online refinement use the 697 test sets to train for 20 epochs and then test the accuracy on the 697 test sets? But in other work(https://arxiv.org/abs/2104.14540), they performed 20 iterations on a single sample of the test set one by one, and loaded the same trained model for the training each sample. Do you have any suggestions

sconlyshootery commented 3 years ago

You can reopen this issue when you have further problem.