add det+rec ckpt prediction pipeline #216
Conversation
HaoyangLee
force-pushed
the
main
branch
2 times, most recently
from
e7ca436 to
f5f92db
Compare
docs/cn/predict_ckpt_cn.md
Outdated
|
|
||
| ```bash | ||
| cd tools/predict/text | ||
| python predict_system.py |
There was a problem hiding this comment.
这里应该要输入config参数？
另外，最好能支持输入待识别图像文件夹路径
There was a problem hiding this comment.
已修改。
- 对齐友商接口，现支持传入raw_data_dir，det_ckpt_path，rec_ckpt_path。
- config文件路径可由det_config_path和rec_config_path参数传入，args parser中有默认值。
- 目前待识别图像文件夹=检测后裁剪图片保存文件夹，可由crop_save_dir参数传入，args parser中有默认值。
add det+rec ckpt prediction pipeline add det_rec ckpt prediction pipeline add det+rec ckpt prediction pipeline add det+rec ckpt prediction pipeline add dbnet+crnn ckpt inference add ckpt predict pytest modify st add predict_system doc update predict_system update doc
There was a problem hiding this comment.
这一步后处理固定写的是4点的文字框。这块后续要考虑对弯曲文字的支持，即根据弯曲的多边形crop图后，padding得到长宽图。
| points = sorted(list(cv2.boxPoints(bounding_box)), key=lambda x: x[0]) | ||
|
|
||
| # index_1, index_2, index_3, index_4 = 0, 1, 2, 3 | ||
| if points[1][1] > points[0][1]: |
There was a problem hiding this comment.
Is there a need for an explicit sorting? cv2.boxPoints() method returns an instance of RotatedRect::points, which in turn returns sorted coordinates is the order of: bottomLeft, topLeft, topRight, bottomRight.
np.roll(..., -1) sets the starting coordinate to topLeft.
| __all__ = ['PredictDataset'] | ||
|
|
||
|
|
||
| class PredictDataset(BaseDataset): |
There was a problem hiding this comment.
I feel like PredictDataset should be a part of DetDataset rather than a separate class. Can set a flag such as prediction to load images only. Perhaps, we can do the same for RecDataset.
| @@ -148,3 +148,41 @@ eval: | |||
| batch_size: 1 # TODO: due to dynamic shape of polygons (num of boxes varies), BS has to be 1 | |||
| drop_remainder: False | |||
| num_workers: 2 | |||
|
|
|||
| predict: | |||
There was a problem hiding this comment.
Because of the separate PredictDataset class, there is many duplicates in the config file. It's easy to make a mistake when there's a lot of repetitions.
add dbnet+crnn ckpt inference
add ckpt predict pytest
Thank you for your contribution to the MindOCR repo.
Before submitting this PR, please make sure:
Motivation
Add detection+recognition end2end ckpt prediction pipeline.
Test Plan
tests/st/test_predict.py will be added later.
Related Issues and PRs
Issue #238