Box2Poly: Memory-Efficient Polygon Prediction of Arbitrarily Shaped and Rotated Text

TL;DR

Box2Poly introduces a memory-efficient, cascade-based polygon prediction method for irregularly shaped and rotated text detection, improving efficiency and maintaining accuracy compared to existing Transformer-based approaches.

Contribution

The paper proposes a novel cascade decoding pipeline inspired by Sparse R-CNN for polygon prediction, significantly reducing memory usage and inference time while preserving detection quality.

Findings

Over 50% memory reduction compared to DPText-DETR

More than 40% faster inference speed than DPText-DETR

Minor performance drop on standard benchmarks

Abstract

Recently, Transformer-based text detection techniques have sought to predict polygons by encoding the coordinates of individual boundary vertices using distinct query features. However, this approach incurs a significant memory overhead and struggles to effectively capture the intricate relationships between vertices belonging to the same instance. Consequently, irregular text layouts often lead to the prediction of outlined vertices, diminishing the quality of results. To address these challenges, we present an innovative approach rooted in Sparse R-CNN: a cascade decoding pipeline for polygon prediction. Our method ensures precision by iteratively refining polygon predictions, considering both the scale and location of preceding results. Leveraging this stabilized regression pipeline, even employing just a single feature vector to guide polygon instance regression yields promising detection results. Simultaneously, the leverage of instance-level feature proposal substantially enhances memory efficiency (>50% less vs. the state-of-the-art method DPText-DETR) and reduces inference speed (>40% less vs. DPText-DETR) with minor performance drop on benchmarks.