CZII CryoET Object Identification Challenge - 4th place solution - yu4u & tattaka

Deposition Overview

The 4th place solution employs heatmap-based particle detection using both 2.5D UNet-like models with 2D image-based backbones and 3D outputs. The solution utilizes weighted mean squared error loss with a balance between positive and negative samples, operates on 32×128×128 voxel patches during training, and converts models to TensorRT format for faster inference. Ensembling combines seven models (four from one contributor and three from another) with test-time augmentation, while post-processing uses non-maximum suppression with a kernel size of 7 and particle-specific thresholds.