method: AntAI-Cognition2020-04-22

Authors: Qingpei Guo, Yudong Liu, Pengcheng Yang, Yonggang Li, Yongtao Wang, Jingdong Chen, Wei Chu

Affiliation: Ant Group & PKU


Description: We are from Ant Group & PKU. Our approach is an ensemble method with three text detection models. The text detection models mainly follow the MaskRCNN framework[1], with different backbones(ResNext101-64x4d[2], CBNet[3], ResNext101-32x32d_wsl[4]) used. GBDT[5] is trained to normalize confidence scores and select quadrilateral boxes with the highest quality from all text detection models' outputs. Multi-scale training and testing are adopted for all basic models. For the training set, we also add ICDAR19 MLT datasets, both training & validation sets are used to get the final result.

[1] He K, Gkioxari G, Dollár P, et al. Mask r-cnn[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2961-2969. [2] Xie S, Girshick R, Dollár P, et al. Aggregated residual transformations for deep neural networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1492-1500. [3] Liu Y, Wang Y, Wang S, et al. Cbnet: A novel composite backbone network architecture for object detection[J]. arXiv preprint arXiv:1909.03625, 2019. [4] Mahajan D, Girshick R, Ramanathan V, et al. Exploring the limits of weakly supervised pretraining[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 181-196. [5] Ke G, Meng Q, Finley T, et al. Lightgbm: A highly efficient gradient boosting decision tree[C]//Advances in neural information processing systems. 2017: 3146-3154.