About Flow module about multiview2novelview

[shaohua0116]

Our flow module is similar to AFN except that the architecture is fully-convolutional (i.e. without FC layers). We do not use DOAFN because it requires 3D supervision (e.g. surface normal), which makes it hard to apply to real-world scenes.

[ehabhelmy82]

Thank you very much for your reply.
One more question, on what basis do you choose the number of iterations used for each data set, for example i noticed that you trained your model as follow:
Synthia, model_checkpoint_path: "model-33001"
Car model_checkpoint_path: "model-140001"
Chair model_checkpoint_path: "model-180001"
KITTI model_checkpoint_path: "model-41001"
Why you stopped kitti at 41001 and car at 180001?
Thanks again

[ehabhelmy82]

???

[shaohua0116]

It takes longer to train models for cars and chairs, partially because we train those models using four source views.

[ehabhelmy82]

I understand this point but my question why you stopped at model 33001 for synthia not 70001 for example?
on what basis do you decide to terminate training at a certain model? why this model specially and why not the same for all data sets?

[shaohua0116]

I stopped the models once the training curves on TensorBoard converge.
Why not looking at validation sets? The main reason is that given the training set is extremely large (considering all the permutations of multiple source/target views), it is less likely a model can simply overfit. Also, I follow the training and test splits proposed in prior work, where no validation sets were provided.

[ehabhelmy82]

Many thanks
Can you kindly give me the command you used to monitor training curve?

[shaohua0116]

You can host TensorBoard under ./train_dir (using tensorboard --logdir . --port 6006) and then you should be able to see this under the IMAGES tab and losses under the SCALARS tab.