This folder contains yeast cell tracking models used in the publication "Enhancing Yeast Cell Tracking with a Time-Symmetric Deep Learning Approach", except for the folder LegacyModels which were not utilized anymore.

The folder Segmentation Models contains Detectron2-based Mask R-CNN implementations, model_larger_set_aug1.pth being the main model trained with the primarily described parameters, while the models in the folder Sample Reduced were trained using reduced number of training samples (SR05 -> 50%, SR02 -> 20%, etc.)
For all of these models, the environment can be configured with config_larger_set.yaml

The folder Tracker Models contains local tracking models with various characteristics, the naming conventions are as follows:
- [Architecture,Backbone]: The used architecture in the Segmentation Models Pytorch implementation (for non-legacy models, this is always DLV3p -> DeepLabV3+), and the used backbone like EFFNetBX -> EfficientNetBX, MNV3min100 -> timm-mobilenetv3_large_100, RNX -> ResNetX. In the paper, only the resnet backbones were explored due to their far superior performance for the task.
- FBtrX: X marks the local tracking range of the model (TR parameter in the paper)
- LargeDataset: The initial models were trained on a much smaller sub-dataset, some of them are included in the legacy folder. In the paper only the models trained on the full dataset were utilized.
- Augm, Augm_Adv1, Augm_Adv2: A buildup of augmentation techniques. In the paper the Augm_Adv2 set of augmentation techniques is described na dutilized only, as it consistantly showed superior performance.
- SRX: Similarly to the segmentation models, it marks the cases where the model was trained on a smaller subset of the full dataset (SR05 -> 50%, SR02 -> 20%, etc.)

The models for synthetic arrows are available at: https://users.itk.ppke.hu/~szage11/General%20tracking/ArrowSynth/ArrowSynth1/
The models for the synthetic amoeboid variants are available at: https://users.itk.ppke.hu/~szage11/General%20tracking/AmobeSynth/
Both folders use similar naming conventions as described before.