OCR Q3
Created by: wanghaoshuang
OCR端到端
1. east检测 (8周)
目标: 对齐baseline 上线模型training数据集:100W级别
1.0 编译环境 @wanghaoshuang
1.1 所需op
- p0: dice-loss (2周*1人)@wanghaoshuang
- p1: 多边形的非最大抑制 (2周*1人) @ocr
- p2: IoU loss (2周*1人)@ocr
- p3: Cos loss (2周*1人)@wanghaoshuang
1.2 模型配置 (1周 * 1人)@ocr
1.3 data reader (1周 * 1人)@ocr
1.4 效果验证 (4周 * 2人)@ocr + @wanghaoshuang
- benchmark training 效果对齐(icdar15)
- 预测库效果对齐(caffe)
2. 识别
2.1 CTC(done)
2.2 Attention (4周 * 1人)@shiwenguo + @wanghaoshuang
- 1d attention效果验证
- 2d attention效果验证
3. 融合 (east检测 + CTC)
3.1 所需ops
- roi_perspective op
3.2 代码框架整合
3.3 合并DataReader
TODO:
- 确认tensorflow&caffe model转fluid model
- 确认具体人力
- 是否可开源
- 确定east检测分支开发环境