BC03 3DoF, 1 Sphere, RGB CNN Baseline (with and without data aug)
Naive method, however I expect it to work well on the translation only case. I'm testing with and without data augmentation, the data augm uses randomized image crops.
Created on May 14|Last edited on May 19
Comment
This is showing expected behavior w.r.t. the MSEs. For training, it should go down. For validation, it should go down and up. Do we see the same thing for the PN++ variants (but note that for these we need to be sure we use the changes ON OR AFTER 05/14/2022 when I made that fix for evaluation).
Overall, the binary success rate is looking good but I would actually argue that we get some better results with the PN++ averaging, but we will want to run this with data augmentation (random cropping). Note: this is with 100x100 input to the CNN, with center-cropping only in training and testing (so later, do random cropping in training only).
Update (05/16/2022): now testing with randomized image crops, instead of just centering crops. Good news: behavior is as expected, we get better test-time generlization with randomized crops while the train MSE is worse at the very beginning (due to need to handle the different augmentations). This is a good sanity check and suggests we can get some gains if we do data augmentation on point clouds. FYI, 1 out of the 5 with random crops didn't do well, fortunately I checked and it is NOT the case that it ran into physics instability, the policy just stagnated. That's quite weird, though, it may be more robust to just remove the best / worst? Averaging is easiest for now.
Section 1
No Data Aug
5
With Data Aug
5
GIFs
With data augmentation, does quite well. Examples after 2 separate runs, at the end of training.
Here's the run that didn't do well for data augmentation at 150 epochs and at 300
No idea what happened here.
Code
Ran after these commits:
Add a comment