We've run an experiment on low data regimes, to see if an augmentation method would allow for a better training with very little data.
Trained chaos dataset with 200 epochs (same config as regular chaos runs). Tested only on in domain data, chaos test set.
Looking at the dice results, our raw method allow for better training at low data regimes, than all other methods. Our method + auglab seems "constrained" by auglab.
However, HD95 shows us more contrasted results. We're still consistently better than synthseg EM. There is big chance the strong auglab augmentation improve robustness which directly impacts HD95.
Worth reporting in the paper?
We've run an experiment on low data regimes, to see if an augmentation method would allow for a better training with very little data.
Trained chaos dataset with 200 epochs (same config as regular chaos runs). Tested only on in domain data, chaos test set.
Looking at the dice results, our raw method allow for better training at low data regimes, than all other methods. Our method + auglab seems "constrained" by auglab.
However, HD95 shows us more contrasted results. We're still consistently better than synthseg EM. There is big chance the strong auglab augmentation improve robustness which directly impacts HD95.
Worth reporting in the paper?