Source-linked AI summary
Phase-Shifting Coder: Predicting Accurate Orientation in Oriented Object Detection
Yi Yu, Feipeng Da
TL;DR
Oriented object detection must predict angles despite boundary discontinuity and square-like ambiguity caused by rotational symmetry. The paper proposes phase-shifting coder (PSC) and dual-frequency PSCD to map rotational periodicities to phase frequencies. Across reported experiments, PSCD improves over CSL and KLD on DOTA, especially for mAP75, supporting competitive performance when high-IoU boxes are required.
Problem
Angular periodicity creates boundary discontinuity, while 90° rotational equivalence creates square-like ambiguity in oriented bounding-box regression.
Method
PSC encodes orientation angles into periodic phases, while PSCD maps different rotational cycles to different frequencies and unwraps the phases.
Results
On DOTA, PSCD outperforms CSL by 1.15 pp in mAP50 and 2.59 pp in mAP75, and outperforms KLD by 2.36 pp in mAP75.
Takeaways & Limitations
PSC and PSCD are expected to provide competitive performance in scenarios requiring high-IoU bounding boxes.
Takeaways & Limitations
PSC's applicability to objects with mixed 360°, 180°, and 90° rotational symmetries remains theoretically proposed and is left for future work.
Abstract
from arXiv · showhide
With the vigorous development of computer vision, oriented object detection has gradually been featured. In this paper, a novel differentiable angle coder named phase-shifting coder (PSC) is proposed to accurately predict the orientation of objects, along with a dual-frequency version (PSCD). By mapping the rotational periodicity of different cycles into the phase of different frequencies, we provide a unified framework for various periodic fuzzy problems caused by rotational symmetry in oriented object detection. Upon such a framework, common problems in oriented object detection such as boundary discontinuity and square-like problems are elegantly solved in a unified form. Visual analysis and experiments on three datasets prove the effectiveness and the potentiality of our approach. When facing scenarios requiring high-quality bounding boxes, the proposed methods are expected to give a competitive performance. The codes are publicly available at https://github.com/open-mmlab/mmrotate.
1. Introduction
Oriented object detection needs angle prediction for applications such as aerial imagery, scene text, and industrial inspection, but angular periodicity creates boundary and square-like ambiguities. The paper adapts phase-shifting coding to unify these problems and evaluates the resulting approach experimentally.
- Oriented bounding boxes are preferable in aerial images, scene text, and industrial inspection, extending detection beyond predominantly horizontal objects.
- Boundary discontinuity makes equivalent angles such as −π/2 and π/2 produce inconsistent regression targets for the same input.
- Square-like ambiguity arises because a square bounding box and its 90° rotation are equivalent despite incurring a high angle-parameter regression loss.
- Boundary and square-like problems are unified as rotationally symmetric periodic fuzzy problems with 180° and 90° cycles, respectively.
- PSC adapts optical phase-shifting by encoding orientation angles into periodic phases, inherently addressing boundary discontinuity; dual-frequency phase-shifting mixes frequencies to address periodic fuzziness.
- The paper presents an integral angle-coding solution, evaluates it through extensive experiments with fair re-testing, and provides publicly available code.
2. Related work
Related work extends horizontal detectors to oriented detection and adds rotation-aware architectures, while specialized methods target boundary discontinuity and square-like ambiguity. Existing approaches trade off theoretical problem handling, simplicity, hyper-parameter sensitivity, and localization accuracy.
- From horizontal to oriented: Representative oriented frameworks include anchor-based RetinaNet, anchor-free FCOS, point-based RepPoint, and high-efficiency YOLO.
- From horizontal to oriented: Horizontal detectors become oriented detectors by adding an output channel for each object's orientation, producing rotated RetinaNet, FCOS, RepPoint, and YOLO variants.
- Rotation-aware architectures: Additional architectures address rotation through region transformations, rotation-equivariant or invariant features, feature alignment, progressive refinement, or oriented proposals and RoI refinement.
- Boundary and square-like problems: Existing specialized methods use smooth losses, angle coders such as CSL, or Gaussian-distribution losses including GWD, KLD, and KFIoU.
- Boundary and square-like problems: CSL is simple and stable but cannot solve the square-like problem and is sensitive to hyper-parameters, whereas GWD and KLD address both problems but yield relatively inaccurate predictions.
- Boundary and square-like problems: GWD and KLD can produce high mAP50 but low mAP75, while the broader field's boundary and square-like problems remain incompletely solved.
3. Method
PSC encodes orientation as periodic phase patterns and decodes them through differentiable phase operations. PSCD adds a second frequency so rotational ambiguities with different cycles can be unified and resolved.
- 3.1. Phase-shifting coder (PSC): A mapping is required because sinusoidal phase has a 2π cycle, whereas a rectangle is identical after π rotation.This aligns the phase representation with the rotational periodicity of oriented bounding boxes.
- 3.1. Phase-shifting coder (PSC): PSC maps an orientation angle to periodic phase-shifting patterns and decodes those patterns back into an angle.The encoding uses Nstep phase-shifting steps, while decoding applies an arctan2-based phase recovery.
- 3.3. General form of mapping: The general mapping uses φ = kθ, with k = 2π/s when objects are symmetric under rotation by s radians.This extension covers triangle-like, pentagon-like, and 360° orientation ambiguities; multiple frequencies handle different periodicities simultaneously.
4. Experiments
Experiments evaluate PSC and PSCD across three datasets, detectors, baselines, and ablations. Results indicate robust high-IoU performance, limited sensitivity to Nstep, and dataset-dependent preference between single- and dual-frequency coding.
- Experimental setup: Experiments use mAP as the major metric and evaluate PSC with FCOS, RetinaNet, and YOLO across DOTA, HRSC, and OCDPCB.Comparisons keep backbone and augmentation identical within each group for fairness.
- Ablation study: PSC has one integer hyper-parameter, Nstep ≥ 3, whose impact on results is limited.The study recommends Nstep = 3 after considering computational complexity; with Nstep = 3, PSC has no adjustable hyper-parameters.
- Ablation study: Dual-frequency PSC is reported to solve boundary discontinuity and square-like problems in a unified way.The authors recommend the dual-frequency strategy for scenarios containing square-like objects.
- Results and analysis: PSCD improves mAP75 over PSC on DOTA, whereas PSC can be preferable on HRSC and OCDPCB, where square-like objects are less common.PSCD raises mAP75 by 3.94 pp on YOLOv5s, 3.60 pp on YOLOv5m, and 0.64 pp on average for FCOS and RetinaNet on DOTA.
- Results and analysis: Across new-dataset circumstances, PSC outperforms KLD by 9.83 pp on average in mAP75 and is reported to obtain high performance more easily than CSL and KLD.On OCDPCB, CSL and KLD show limited or negative results without dataset-specific hyper-parameter tuning.
5. Conclusions
The paper proposes PSC and PSCD as phase-based angle coders for oriented object detection, addressing boundary discontinuity and square-like problems. Experiments report improved high-IoU performance, while the preferred coder depends on the prevalence of square-like objects.
- 5. Conclusions: PSC encodes orientation angles into periodic phases, while PSCD uses different phase frequencies to address both boundary discontinuity and square-like problems.PSCD extends PSC by adding a frequency for another rotational periodicity.
- 5. Conclusions: PSC and PSCD improve mAP50 and mAP75 over CSL, while PSCD exceeds KLD by an average of 2.36 pp in mAP75 on DOTA.The paper reports these comparisons across oriented detection experiments.
- 5. Conclusions: On DOTA, PSCD improves mAP75 over PSC by 3.94 pp with YOLOv5s and 3.60 pp with YOLOv5m.The dual-frequency strategy is especially beneficial on this dataset.
- 5. Conclusions: On HRSC and OCDPCB, the dual-frequency strategy has a negative effect, so PSC may be preferable when few square-like objects are present.The reported dataset-dependent difference links PSCD’s benefit to the presence of square-like objects.
- 5. Conclusions: PSC and PSCD avoid dataset-dependent hyper-parameter tuning, supporting more comfortable application to new scenarios.This contrasts with existing approaches that require different parameters across datasets.
- 5. Conclusions: Mixed scenes containing objects with 360°, 180°, and 90° rotational requirements remain a future-work setting for PSC.The paper states that PSC is theoretically applicable to this extended square-like problem but leaves further exploration for future work.