Source code of the 7th place solution (gold medal) for Sartorius - Cell Instance Segmentation from team Train4Ever.

1. About the challenge

• The goal of Sartorius - Cell Instance Segmentation challenge on Kaggle is to segment individual neuronal cells in microscopic images.
• Note that we need to distinguish among different cells as separated instances, you can refer to this link if not familiar with instance segmentation task. This is fundamentally different from semantic segmentation where only a binary mask is predicted to mark whether a pixel belongs to a foreground class or background.

2. Solution overview

Detail explanation can be found in this Kaggle discussion: https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/298002

The ensembling technique used to combine 2 final segmentation model was conducted as this flow:

3. How to run the code

We have tried many different models such as PointRend, MaskRCNN Swin Transformer, CellPose, GCNet, etc. However the final solution is the ensembling prediction of 2 MaskRCNN ResNeSt200 model, followed by a 2nd-level catboost post-processing model. Thus, to keep everything simple, we only provide the training, evaluation and inference code of MaskRCNN ResNeSt200 and the 2nd-level catboost model.

For a quick demo on colab: https://colab.research.google.com/drive/1X2gI5jW5CrjY_gH3Cp5-oTG9c8pY47Tt?usp=sharing

3.1. Install dependecies

    bash install_libs.sh

3.2. Download pretrained models

    bash download_models.sh

3.3. Download data

    bash download_data.sh

3.4. Run end-to-end inference on an image

    python src/infer_cis.py --image [path to image] \
                            --weights [path to weights separated by space, enclosed in quote marks] \
                            --2nd_level_model [path to 2nd-level model saved in pickle file] \
                            --2nd_level_features [path to 2nd-level features saved in csv file] \
    
    Default: python src/infer_cis.py --image demo_images/7ae19de7bc2a.png \
                                    --weights 'models/pretrained_models/pseudo_round1_model.pth models/pretrained_models/pseudo_round2_model.pth' \
                                    --2nd_level_model 2nd_level_model/catboost.pkl \
                                    --2nd 2nd_level_features 2nd_level_model/features.csv

Output image will be saved in demo_outputs/ folder. It should look like this:

3.5. Train instance segmentation models

    python src/train_cis.py --image_dir [path to folder containing training images] \
                            --batch_size [training batch size] \
                            --annotation_dir [path to folder containing json annotations] \
                            --out_dir [path to output folder where trained weights will be saved]

Round-1 pseudo training:

    python src/train_cis.py --image_dir data/images \
                            --batch_size 1 \
                            --annotation_dir data/annotations_semi_supervised_round1 \
                            --out_dir models/maskrcnn_ResNeSt200_pseudo_round1_fold0 

Round-2 pseudo training:

    python src/train_cis.py --image_dir data/images \
                            --batch_size 1 \
                            --annotation_dir data/annotations_semi_supervised_round2 \
                            --out_dir models/maskrcnn_ResNeSt200_pseudo_round2_fold0 

3.6. Evaluate single segmentation model

    python src/evaluate_single_model.py --image_dir [path to folder containing validation images] \
                                        --annotation_dir [path to folder containing json annotations] \
                                        --weight [path to trained weight pth file]

    Default: python src/evaluate_single_model.py --image_dir data/images \
                                                --annotation_dir data/annotations_semi_supervised_round2 \
                                                --weight models/pretrained_models/pseudo_round2_model.pth

3.7. Evaluate model ensembling

    python src/evaluate_ensemble.py --image_dir [path to folder containing validation images] \
                                    --annotation_dir [path to folder containing json annotations] \
                                    --weights [path to weights separated by space, enclosed in quote marks]

    Default: python src/evaluate_ensemble.py --image_dir data/images \
                                            --annotation_dir data/annotations_semi_supervised_round2 \
                                            --weight 'models/pretrained_models/pseudo_round1_model.pth models/pretrained_models/pseudo_round2_model.pth'

3.8. Training second 2nd-level model (catboost):

3.8.1. Generate predicted instances on validation set:

    python src/prepare_for_2nd_level_model.py --image_dir [path to folder containing validation images] \
                                            --annotation_dir [path to folder containing json annotations] \
                                            --weights [path to weights separated by space, enclosed in quote marks]

    Default: python src/prepare_for_2nd_level_model.py \
                    --image_dir data/images \
                    --annotation_dir data/annotations_semi_supervised_round2 \
                    --weights 'models/pretrained_models/pseudo_round1_model.pth models/pretrained_models/pseudo_round2_model.pth'

Pickle output files will be saved inside folder: data_for_2nd_level_model/

3.8.1. Train catboost model

    python src/train_2nd_level_catboost.py

Trained catboost model and engineered features are saved inside folder: 2nd_level_model/