Abstract
Detecting oriented text in natural scenes presents a unique challenge: classification requires features that are invariant to rotation, while regression requires features that are highly sensitive to it. The Rotation-sensitive Regression Detector (RRD) addresses this fundamental conflict by decoupling these tasks into two specialized branches. By utilizing rotation-invariant features for classification and rotation-sensitive features for regression, RRD achieves state-of-the-art accuracy in detecting long, tilted text lines across multiple benchmarks.
Introduction
Text detection in "the wild" is notoriously difficult because text can appear at any angle. Standard object detectors typically use horizontal bounding boxes, which are a poor fit for long, thin, diagonal text strings—they include too much background noise and make precise localization nearly impossible. While oriented detectors have been developed to solve this, most rely on a shared feature map for two distinct tasks: classification (determining if a region contains text) and regression (predicting the exact coordinates of the oriented box).
This paper identifies a "feature incompatibility" in this shared approach. Classification benefits from rotation invariance, where features remain stable regardless of tilt. Conversely, regression requires rotation sensitivity to accurately map coordinates to specific orientations. Forcing a single feature map to perform both tasks leads to a performance compromise.
Methodology: The RRD Architecture
RRD is built upon the Single Shot MultiBox Detector (SSD) architecture but introduces a novel "Rotation-Sensitive Regression" (RSR) module. The key technical innovation lies in how features are extracted and utilized.
Oriented Response Convolution (ORN)
Instead of standard convolutions, RRD uses filters that are actively rotated during the forward pass. This produces a rotation-sensitive feature map where different channels represent the text's response at different discrete angles.
The Two-Branch Strategy
To resolve the feature conflict, RRD splits the detection head into two specialized branches:
- Regression Branch: Uses raw rotation-sensitive features directly, allowing the network to "feel" the orientation for precise coordinate prediction.
- Classification Branch: Applies Oriented Response Pooling, which extracts the maximum response across all orientation channels. This creates a rotation-invariant feature map robust for identifying text presence.
Handling Extreme Aspect Ratios
Standard square kernels are ill-suited for text lines that may be significantly longer than they are tall. RRD replaces standard kernels with Inception-style blocks containing multi-scale, asymmetric filters (e.g., $1 \times 9$ and $9 \times 1$). This provides a flexible receptive field that better covers elongated text strings.
Experimental Results
The authors evaluated RRD on a wide array of challenging benchmarks, including RCTW-17, MSRA-TD500, ICDAR 2015, and COCO-Text. RRD achieved state-of-the-art (SOTA) results across almost all tested datasets at the time of publication.
Key highlights include:
- Precision and Localization: On COCO-Text, RRD showed a massive lead when evaluation criteria were made stricter (IoU threshold of 0.75), proving the superior accuracy of its regression branch.
- Generality: On the HRSC2016 ship detection dataset, RRD achieved an mAP of 84.3, an 8.6-point jump over previous specialized ship detectors. This demonstrates that the architecture generalizes to any oriented object detection task beyond just text.
Limitations and Future Work
Despite its strengths, RRD faces challenges with large character spacing, where it may detect individual fragments instead of a single bounding box. It also occasionally struggles with vertical text ambiguity, interpreting vertical lines as stacked horizontal boxes. Future directions include incorporating higher-level semantic understanding and spatial analysis to handle these complex layouts, particularly in aerial imagery.