- Task 1 - Text Localization - Method: Unet Segmentation and Watershed
- Method info
- Samples list
- Per sample details
method: Unet Segmentation and Watershed2019-04-22
Authors: Roberto Lotufo, Israel Campiotti, Ramon Pires, Rubens Machado, Luis Serrano, Giovanni Garuffi
Description: Crop based on horizontal and vertical projections followed by a volume 1D morphological filters on the projections. DynamicUnet built with fastai library with two segmentations maps: the bounding box and the marker for the bounding box. Watershed pos-processing from markers to bounding boxes to disconnect vertical overlapping. Connected component analysis to generate the final bounding boxes coordinates. Data augmentation with translation, rotation, cropping and additive gaussian noise. Loss function based on dice and binary cross entropy.