Dense Label Encoding for Boundary Discontinuity Free Rotation Detection

Abstract

Rotation detection serves as a fundamental building block in many visual applications involving aerial image, scene text, and face etc. Differing from the dominant regression-based approaches for orientation estimation, this paper explores a relatively less-studied methodology based on classification. The hope is to inherently dismiss the boundary discontinuity issue as encountered by the regression-based detectors. We propose new techniques to push its frontier in two aspects: i) new encoding mechanism: the design of two Densely Coded Labels (DCL) for angle classification, to replace the Sparsely Coded Label (SCL) in existing classification-based detectors, leading to three times training speed increase as empirically observed across benchmarks, further with notable improvement in detection accuracy; ii) loss re-weighting: we propose Angle Distance and Aspect Ratio Sensitive Weighting (ADARSW), which improves the detection accuracy especially for square-like objects, by making DCL-based detectors sensitive to angular distance and object's aspect ratio. Extensive experiments and visual analysis on largescale public datasets for aerial images i.e. DOTA, UCAS-AOD, HRSC2016, as well as scene text dataset ICDAR2015 and MLT, show the effectiveness of our approach. The source code is available at https://github.com/ Thinklab-SJTU/DCL_RetinaNet_Tensorflow and is also integrated in our open source rotation detection benchmark: https://github.com/yangxue0827/ RotationDetection.

Optimization for Arbitrary-Oriented Object Detection via Representation Invariance Loss

“…As shown in Tab. 4, RIDet outperforms other single-stage detectors and even some recent advanced two-stage detectors such as DCL [32], RoI Transformer [2]. RIDet-O and RIDet-O achieve the mAP of 89.10%, and 89.63%, respectively.…”

“…However, the weight of skew IoU in IoU-Smooth Loss [39] is not differentiable, and MR Loss [23] only considers limited redundant representations. DCL [32] converts the angle regression into fine-grained angle classification to eliminate the problems caused by the angle boundary, and thus achieving the mAP of 89.46%. But it brings additional computational overhead and cannot be directly used to other detectors.…”

“…Some previous work summarized the issues as the boundary problem [39,36] in OBB representation, and suggested that this will cause a discontinuous regression loss and hinder the convergence of network training. To solve the issues, Circular Smooth Label [36] and Densely Coded Labels [32] transform the angle regression task into a classification task to avoid the discontinuity of regression loss. SCRDet [39] adopts IoU-Smooth-L 1 loss to alleviate the sharp increase of loss caused by angle boundary.…”

Learning Polar Encodings for Arbitrary-Oriented Ship Detection in SAR Images

“…3(b), in the boundary case, the edges of the predicted OBB and the ground truth correctly correspond to each other, but the angle suffers discontinuity because of the POA. In addition, the performance of the angle prediction-based methods is sensitive to angle prediction errors [49], [56]. As shown in Fig.…”

“…The fore-mentioned OBB-based methods encounter the boundary discontinuity problem in varying degrees. The causes of the problem can be attributed to two sides, the periodicity of angle (POA) and the exchangeability of edge (EOE) [45], [48], [56]. The boundary discontinuity problem leads to mismatching between annotations and predictions during the training stage, causing performance degradation.…”