The singular defects (or high-norm tokens) [1] may be related to attention sinks. It is interesting that the direction of all high-norm tokens share the same direction. Maybe the theory behind is not very complex and the issue can be fixed cleverly during training.
For more text classification baselines (CRNN, NRTR, RubustScanner, SAR, SegOCR), checkout https://github.com/open-mmlab/mmocr They are reproducible, customizable.
We also expored ways of presenting hundreds [1] of AI models in a better way. Recently we updated the model zoo so that it is arranged according to benchmark datasets [2] and papers [3]. Do you think it satisfies the needs?
One of the maintainer here. We just now open-sourced a new ocr library. It contains 5 text detection algorithm, 4 text recognition algorithms, and 1 key information extraction algorithm (details see [1]). It is used internally for some time. We are exicted to join the OpenMMLab [2] ecosystem, and we hope you could try it, and let issues and prs coming!
Checkout the openmmlab project [1], where messy research codes in computer vision are rewritten an reorganized into a coherent whole. If there are more researchers join this, then not only research are fully reproducible, much more accessible to everyone, but also be compared fairly. (I'm the maintainer of mmpose [2], mmaction2 [3], and mmediting [4])
The single RGB stream top1 goes up to 73.48% with resnet50, and up to 74.71% equipped with non-local. Both are much higher than the original paper with two-streams.
MMPose is an open-source toolbox for pose estimation based on PyTorch. It is a part of the OpenMMLab project. In V0.7.0, it adds support for the 3D human shape recovery task, adds support for WholeBody human pose estimation task, and the 2D hand keypoint estimation task.
[1] https://openreview.net/pdf?id=4yBnUokU2v