aengusng8/Sartorius-Cell-Instance-Segmentation

Instance segmentation: two training stages with transfer learning, two inference stages with EfficientNetV2 and Mask R-CNN R101-FPN, ensemble masks with Weighted Boxes Fusion, heavily rotated Mosaic w/o artifact or tiny bounding boxes, Test time augmentation, five-fold cross-validation, Detectron2 with Albumentations, CIoU loss andGIoU loss.

Sartorius Cell Instance Segmentation

Summary

• Model 1 (Classifier): EfficientNetV2
  • train with train_semi_supervised and Sartorious dataset
  • result: 0.0 losses and 100% accuracy on training, validation sets
• Model 2 (Instance Segmentation): Mask R-CNN (R101-FPN)
  • Training stage 1: train with eight-class LIVECell dataset
  • Training stage 2: train with three-class Sartorious dataset with pretrained weights from stage 1
• Inference: use EfficientNetV2 classifier to predict class which is used to refine final prediction of Mask R-CNN
• 5 folds cross-validation
• Ensembling masks from different models with customed Weighted Boxes Fusion

What didn't work

• Heavy augmentation:
  • 2 x customed Mosaic data augmentation (deleted tiny bouding boxes)
    • RandomRotate + CenterCrop (w/o artifact) || Transpose (Reflection) + RandomCrop
    • RandomCrop + random x_center and y_center
  • MixUp
  • AdditiveGaussianNoise, GaussNoise, MotionBlur, MedianBlur, Blur, CLAHE, Sharpen, Emboss, RandomBrightnessContrast
• Test time augmentation (HorizontalFlip, VerticalFlip)
• CIoU loss and GIoU loss

Miscellaneous:

annotations_train.json:

• segmentation dict: represents the per-pixel segmentation mask in COCO’s compressed RLE format. The dict should have keys “size” and “counts”.
• You can convert a uint8 segmentation mask of 0s and 1s into such dict by pycocotools.mask.encode(np.asarray(mask, order="F")).
• cfg.INPUT.MASK_FORMAT must be set to bitmask if using the default data loader with such format.