Wrong bounding boxes of results, but recognitions are correct?

[doanthuyennt]

Mango prediction results are weird? Is it related to https://github.com/hikopensource/DAVAR-Lab-OCR/blob/main/davarocr/davar_spotting/models/spotters/mango.py#L276 ? Or maybe not? How can i get correct bounding box (detection)?

[qiaoliang6]

MANGO is a pure single-staged end-to-end framework，where the recognizor is predicted based on the attention of the global feature map. The detector here is only used to do coarse location (use shrinked text). If you want to obtain more accurate text locations, you may adopt any text detection head in this branch. You may refer to the paper for more details.