Qing

I think it's because PyTorch is slow on the M1 chip, here is some test result on my MacBook Pro (16-inch) Intel i9: - Original(1920 * 1000): 44s - Resize(1280...

Complete remove landing page: https://github.com/Sanster/lama-cleaner/releases/tag/0.22.0 https://user-images.githubusercontent.com/3998421/194713446-c22e8320-1f12-4910-814b-ae2b0f1a3ade.mov https://user-images.githubusercontent.com/3998421/194713757-21b637a1-fc5d-408a-987b-337627a3f33b.mov

There are also some images have wrong labels, like img_3591.jpg  img_5737.jpg and img_5741.jpg update: These images are wrong labeled in original MLT17 training dataset

I am trying to clean the data and recreate the anchor labels from MLT17 according to the `minAreaRect` of a text line. Not sure whether the training result will be...

After recreate the ground truth labels and make several changes (see https://github.com/Sanster/tf_ctpn/commit/dc533e030e5431212c1d4dbca0bcd7e594a8a368 and https://github.com/Sanster/tf_ctpn/commit/7ae3d50d72bbdccb16f00987a5edb97659d6fbf2), I got better result on ICDAR13: | Net | Dataset | Recall | Precision | Hmean...

@interxuxing Maybe - **No side-refinement part** - **Different way from Conv5 to BLSTM see https://github.com/eragonruan/text-detection-ctpn/issues/193** - The training data is different - Use `adam`. Origin CTPN use SGD - ...

@interxuxing I think it worth a try

- Cleaned data MLT17(latin+chinese) + ICDAR13: [google drive](https://drive.google.com/file/d/1S9K9NKkA0RYlBswCfyUI0dv_fI4r5bcX/view) - Code for split text line by `minAreaRect`: [mlt17_to_voc.py](https://github.com/Sanster/tf_ctpn/blob/master/tools/mlt17_to_voc.py)

@saicoco The result of "origin CTPN" is from ICDAR13 result page

@Wangweilai1 I think vertical words(not suitable for CTPN), very small words(can't recognize by human) should be negative examples, or we can create a ignore mask like in [EAST](https://github.com/argman/EAST/blob/ab97939783901b7e22ff55e151964e159d1627b9/icdar.py#L485). Not sure...