I3CL: Intra- and Inter-Instance Collaborative Learning for Arbitrary-Shaped Scene Text Detection

Existing methods for arbitrary-shaped text detection in natural scenes face two critical issues, i.e., (1) fracture detections at the gaps a instance; and (2) inaccurate of instances with diverse background context. To address these we propose novel method named Intra- Inter-Instance Collaborative Learning (I3CL). Specifically, to first issue, design an effective convolutional module multiple receptive fields, which is able collaboratively learn better character gap feature representations local long ranges inside instance. second devise instance-based transformer exploit dependencies between different global context semantic from shared background, are more discriminative representation. In this way, I3CL can effectively intra- inter-instance together unified end-to-end trainable framework. Besides, make full use unlabeled data, semi-supervised learning leverage pseudo labels via ensemble strategy. Without bells whistles, experimental results show that proposed sets new state-of-the-art on three challenging public benchmarks, F-measure 77.5% ArT, 86.9% Total-Text, 86.4% CTW-1500. Notably, our ResNeSt-101 backbone ranked 1st place ArT leaderboard. Code available www.github.com/ViTAE-Transformer/ViTAE-Transformer-Scene-Text-Detection .