CZII CryoET Object Identification Challenge - 7th place solution - kobakos

Deposition Overview

The 7th place solution employs an ensemble of three model soups based on 3D U-Net architectures with ResNet50d and EfficientNetV2-M backbones, training them to predict per-class Gaussian heatmaps using heavily weighted BCE loss with extraordinarily high positive weights to increase recall. The models are first pretrained on simulated data before fine-tuning on experimental data with extensive augmentations including Mixup, Cutmix, RandomFlip, and rotations on the xy plane. Local maxima detection is optimized through Gaussian smoothing before maxpool application and weighted box fusion, while ensemble predictions are combined at the logit level rather than probability level, with final thresholds applied in logit space.

Annotation Methods Summary