CZII CryoET Object Identification Challenge - 1st place solution - Daddies

Deposition Overview

The 1st place solution employs an ensemble approach combining segmentation models (3D UNets with ResNet & B3 encoders) and object detection models (SegResNet and DynUnet backbones) from MONAI. Segmentation uses weighted CrossEntropy loss (256:1 positive:negative weighting), while detection implements a modified PP-Yolo loss with IoU-based point-point similarity metrics. Models are trained on 96×96×96 patches with inference on larger volumes, and both approaches are merged through a novel scaling technique that aligns feature map distributions before object detection post-processing. Performance is optimized by converting models to TensorRT, achieving a 200% speedup and enabling parallel inference on two T4 GPUs.

Annotation Methods Summary