CZII CryoET Object Identification Challenge - 9th place solution - Avengers

Deposition Overview

For the 9th place solution, a 3D ConvNeXt-like segmentation model is employed with modifications to the encoder including smaller 2x2 stem and 3x3 kernel size in convolution blocks to better detect small particles. Binary cross-entropy loss is used for training with basic rot90 augmentations, and the final submission ensembles multiple models with DBSCAN clustering applied to refine particle centroids. A key distinction of this approach is the custom adjustment of ground truth mask sizes for different particle types with factors of 0.5 for smaller particles and 0.33 for larger ones, which significantly improved performance from 0.62 to 0.70-0.77 on the public leaderboard.

Annotation Methods Summary