Bridging the Gap Between End-to-End and Two-Step Text Spotting.

Modularity plays a crucial role in the development and maintenance of complexsystems. While end-to-end text spotting efficiently mitigates the issues oferror accumulation and sub-optimal performance seen in traditional two-stepmethodologies, the two-step methods continue to be favored in many competitionsand practical settings due to their superior modularity. In this paper, weintroduce Bridging Text Spotting, a novel approach that resolves the erroraccumulation and suboptimal performance issues in two-step methods whileretaining modularity. To achieve this, we adopt a well-trained detector andrecognizer that are developed and trained independently and then lock theirparameters to preserve their already acquired capabilities. Subsequently, weintroduce a Bridge that connects the locked detector and recognizer through azero-initialized neural network. This zero-initialized neural network,initialized with weights set to zeros, ensures seamless integration of thelarge receptive field features in detection into the locked recognizer.Furthermore, since the fixed detector and recognizer cannot naturally acquireend-to-end optimization features, we adopt the Adapter to facilitate theirefficient learning of these features. We demonstrate the effectiveness of theproposed method through extensive experiments: Connecting the latest detectorand recognizer through Bridging Text Spotting, we achieved an accuracy of 83.3on Total-Text, 69.8at https://github.com/mxin262/Bridging-Text-Spotting.