# 15th place solution

Thank you for organizers hosting the competition. I also appreciate my team mates cooperation.
Summary
Our solution has nothing special, it's just ensemble of 4 EfficientNetV2 with different image sizes.
2x l with 768x768 image size (different seeds)
1x l with 1024x1024 image size
1x s with 1536x1536 image size
First 3 models give similar score around 0.690 ~ 0.692 for
validation
folder images and the last gives 0.685. Ensemble score is 0.6936.
What worked
Augmentations
HV Flip
Rotate
Random scale then crop
TTA
HV Flip
Rotate
set
"drop_path_rate": 0.2
and
"drop_rate": 0.2
Adding time steps 3 and 5 as pseudo labels
Focal loss
But I guess improvements from above things are pretty small compared with addressing label misalignment and using individual human annotated mask as new labels as mentioned other discussions (I'm plan to do experiment later).
What not worked
Other augmentations
ColorJitter
HueSaturationValue
RandomBrightnessContrast
RandomGamma
RandomFog
RandomShadow
CoarseDropout
MixUp
Adding more time steps as pseudo labels
Adding pseudo label rounds
I'm really focused on training 2.5D and 3D models, but nothing worked at all.
Just 3D model (e.g. Resnet3D)
3D encoder + 2.5D decoder
Input other time frames predictions as additional channels
Stack 2 Unets, first predicts segmentation masks for multiple time frames, second takes those masks as input by stacking channel dimensions