Bad images in training
Closed this issue · 7 comments
While playing around with the sm5 dataset, I noticed some of them are badly rendered.
Not sure if this will pose any problem for training, just wanted to point this out.
Are you using the dataset you created yourself?
Yes. I've seen several sequences with more than 2 objects in it, like the one I posted here. The image in row 2 column 2 and row 3 column 1 clearly shows 2 different objects (one piece with 3 blocks and another piece with 2 blocks (yellow, blue)).
I've created my own dataloader based on https://github.com/l3robot/gqn_datasets_translator and I've observed the same kind of poor quality samples. Might such samples cause exploding gradients in the network?
I believe the problem with these images is that the object is not at the center, but the rendering is actually correct.
Yea I agree, but having some non-centered data(as well as multiple objects in an image) is not representative of the rest of the data on which the network is trying to learn so I would guess it could result in large errors. I'm not sure whether it becomes a problem or not if the amount of such images is small.
I was wondering if you could detect this kind of stuff by looking at the pose? for example if one fit a sphere to the translations of normal ones, the center is at (0, 0, 0) while that of a bad one would be off by a bit.
It looks to only be a few examples at this point, so I conclude it is a problem from the side of DeepMind rather than the converter.