When I wrote pipelines involving mmdetection models and mmocr models for object detection/instance segmentation and text recognition respectively it was very conventient to have the online crop function used with the format x1 y1 x2 y2 x3 y3 x4 y4 , so as not to have to crop the detected boxes and write to disk before doing the text recognition. I haven't found a way to implement this with mmocr 1.0, where online crop is removed. TextRecogCropPacker seemes to write the cropped images to disk, and I want to avoid this, doing it online instead during training and testing. Is this a planned feature, or is it already possible to do this in mmocr 1.0 in some way I haven't figured out?
gaotongxiao
commented
1 year ago
What is the feature?
When I wrote pipelines involving mmdetection models and mmocr models for object detection/instance segmentation and text recognition respectively it was very conventient to have the online crop function used with the format x1 y1 x2 y2 x3 y3 x4 y4 , so as not to have to crop the detected boxes and write to disk before doing the text recognition. I haven't found a way to implement this with mmocr 1.0, where online crop is removed. TextRecogCropPacker seemes to write the cropped images to disk, and I want to avoid this, doing it online instead during training and testing. Is this a planned feature, or is it already possible to do this in mmocr 1.0 in some way I haven't figured out?
Any other context?
No response