# [TextBoxes++: A Single-Shot Oriented Scene Text Detector](https://arxiv.org/abs/1801.02765)
They introduce a new model to detect arbitrary oriented text
and propose a novel score to refine combination of detection and recognition

### Proposed Method :

- The proposed model uses arbitary oriented quadrilaterals and rotated rectangles which are obtained from regressing on default anchor boxes for detection.

- The architecture contains several convolutional and pooling layers followed by textbox layers which are then fed to undergo non maximum suppression.
- The model uses default boxes of various aspect ratios to capture text efficiently and also uses vertical offset to improve text detection in vertical manner
- Both the representations are optimised using the dault boxes and the loss used combines a confidence score and location loss which are obtained using L1 and softmax.
- X is match indication matrix, c is confidence, l is location and g is ground truth.
- To remove false detections due to texture similarity the model changes the ratios of negatives and positives to 3:1 and then 6:1 .
- For data augmentation they propose a new method based on Jacquard called object coverage given by
- During testing they use NMS method in a 2 step fashion first on minimum horizontal textboxes then on quadrilaterals and rotated rectangles, this is done to save time as first step reduces a lot of false candidates.
- The model is refined using end to end recognition by using CRNN as the text recognizer with a given lexicon , the outputs of detection and recognition are combined to a threshold score S used to train the model which is given by:

### Experiments :
- The datasets used are SYnth-Text,IC13,SVT,IC15 andCOCO-Text.
- The metrics used are precision, recall and f-score(f-score used for end-to-end recognition and word spotting also)
- The model achieves SOTA on all the tasks(word spotting,end-to-end recognition and text localisation)
- The model fails to perform well when there is object occlusion, large spacing and for few instances of vertical text.