Source-linked AI summary
TDFNet: Tri-projection Deformable Fusion Network for Panoramic Salient Object Detection
Qiangqiang Zhou, Jiacong Yu, Jiawei Xu, Yong Chen, Xin Huang, Ping Li
TL;DR
Panoramic salient object detection is limited by geometric distortions from ERP polar stretching and CMP cube-face discontinuities. TDFNet combines ERP, CMP, and Tangent Projection through cross-projection deformable attention and latitude-guided fusion, with experiments reporting consistent improvements over existing state-of-the-art methods. The paper’s design is scoped by coordinate and geometric assumptions in CDA and LGF.
Problem
Projection-based panoramic salient object detection lacks a representation that simultaneously preserves global continuity, local geometric fidelity, and boundary consistency.
Method
TDFNet uses ERP, CMP, and Tangent Projection branches with CDA for geometry-aware cross-projection aggregation and LGF for latitude-guided fusion and tangent-reference refinement.
Results
TDFNet achieves consistent quantitative and qualitative improvements over existing state-of-the-art methods on four widely used PSOD benchmarks.
Takeaways & Limitations
The tri-projection design integrates global structural continuity, local geometric fidelity, and fine-grained boundary information for PSOD.
Takeaways & Limitations
CDA is limited to ERP and CMP because tangent features do not share their unified latitude–longitude coordinate system, while LGF treats latitude as a geometric reliability prior rather than semantic confidence.
Abstract
from arXiv · showhide
Recent years have witnessed the growing potential of panoramic salient object detection in robotic vision, virtual reality, and related applications. However, projecting spherical scenes onto 2D planes inevitably introduces geometric distortions, which fundamentally limit the effectiveness of existing projection-based methods. Specifically, Equirectangular Projection (ERP) suffers from severe polar stretching distortions, while cube map projection introduces discontinuities across cube-face boundaries, resulting in degraded feature discriminability and compromised geometric consistency. To address these limitations, we propose TDFNet, the first Tri-projection Deformable Fusion Network for panoramic salient object detection, exploiting complementary projection representations to alleviate geometric distortions and improve detection performance.Specifically, we design a cross-projection deformable attention (CDA) module that leverages spatial correspondences between different projections to construct geometry-aware sampling locations, guiding deformable attention for cross-projection contextual aggregation and enhancing robustness against projection-induced deformations. Furthermore, we introduce a latitude-guided fusion module, which utilizes spherical latitude priors to construct geometric confidence weights for adaptively balancing ERP and CMP features. Meanwhile, LGF incorporates distortion-reduced semantic references from Tangent Projection to achieve cross-projection feature refinement and spatial alignment.By constructing a three-branch encoding architecture based on ERP, CMP, and Tangent Projection, TDFNet simultaneously preserves global spatial continuity, local geometric details, and fine-grained boundary information.
1 Introduction
Panoramic salient object detection benefits from complete spherical scene coverage but is hindered by projection-induced geometric distortions. TDFNet addresses these limitations by combining ERP, CMP, and Tangent Projection with deformable cross-projection fusion.
- Panoramic salient object detection supports robotic vision, virtual reality, and omnidirectional perception by providing complete scene coverage.
- ERP preserves spherical topology but suffers polar stretching, whereas CMP reduces polar deformation while introducing cube-face discontinuities.
- Existing ERP and CMP methods lack a unified representation that preserves global continuity, local geometric fidelity, and boundary consistency simultaneously.
- TDFNet introduces a three-branch ERP, CMP, and Tangent Projection architecture to integrate complementary global structures and low-distortion local geometric details.
- CDA uses cross-projection spatial correspondences for geometry-aware deformable attention, while LGF uses latitude priors and tangent references for adaptive feature alignment and fusion.
- Experiments on four widely used PSOD benchmarks report consistent quantitative and qualitative improvements over existing state-of-the-art methods.
2 Related Work
Related work in salient object detection spans planar, multimodal, and panoramic settings, but panoramic methods remain constrained by trade-offs between global continuity and local geometric distortion. These limitations motivate dedicated multi-projection architectures.
- Deep CNN and Transformer methods capture long-range dependencies and global context for salient object detection in planar images.
- Multimodal salient object detectors combine RGB with depth or thermal infrared cues to address unimodal limitations.
- These projection trade-offs make a dedicated multi-projection architecture necessary for preserving distortion-free local fidelity and global structural continuity.
- Panoramic salient object detection requires binary masks and well-defined object boundaries, unlike panoramic saliency prediction’s continuous fixation density maps.
- ERP provides complete spherical coverage but stretches near the poles, while CMP reduces local distortion but creates discontinuities across cube-face boundaries.
3 Methods
TDFNet uses ERP, CMP, and Tangent Projection branches to preserve complementary global, local geometric, and boundary information. CDA establishes cross-projection geometric priors for ERP and CMP, while LGF progressively fuses the enhanced representations with tangent features.
- Tri-branch Encoder: The tri-branch encoder combines ERP, CMP, and Tangent Projection to capture global continuity, locally preserved geometry, and fine-grained local semantics.ERP retains panoramic continuity, CMP reduces polar stretching through perspective faces, and tangent views provide distortion-free local cues.
- Tri-branch Encoder: The tangent branch samples 18 local tangent viewports arranged across four latitude rows for distortion-free local representation.The rows contain 3, 6, 6, and 3 longitudinal centers, respectively.
- Cross-projection Deformable Attention: CDA constructs hybrid reference points from ERP–CMP spatial correspondences, constraining deformable sampling with cross-projection geometric neighborhoods.This replaces reliance on the regular feature grid alone and supplies a geometry-aware sampling prior for both projection branches.
- Cross-projection Deformable Attention: Deformable Transformer aggregation adaptively samples flattened multi-scale ERP and CMP features using learned offsets and attention weights.CDA enhances ERP and CMP separately while allowing each branch to exploit geometrically complementary neighborhoods and retain projection-specific characteristics.
- Latitude-Guided Fusion: LGF progressively integrates the CDA-enhanced ERP and CMP features with tangent representations through geometry-aware fusion and tangent-guided refinement.The resulting fused representation is passed to the decoder for saliency prediction.
4 EXPERIMENTS
TDFNet is evaluated across four panoramic salient object detection benchmarks against state-of-the-art methods, showing consistently strong quantitative and qualitative performance. Ablations indicate that combining ERP, CMP, and Tangent Projection improves robustness to projection distortions, boundary artifacts, and complex scenes.
- Experimental setup: TDFNet is evaluated on 360-SOD, 360-SSOD, F-360iSOD, and ODI-SOD using four standard saliency metrics and comparisons with 19 existing methods.The evaluation includes panoramic and 2D salient object detection methods.
- Quantitative evaluation: TDFNet achieves consistently superior performance across the four panoramic salient object detection benchmarks.The reported cross-dataset outcome is framed as evidence of effectiveness under projection distortion and complex panoramic structures.
- Quantitative evaluation: On 360-SOD, TDFNet obtains S-measure 0.883, MAE 0.015, mean E-measure 0.927, max E-measure 0.932, mean F-measure 0.835, and max F-measure 0.847 against SCFA.MAE decreases by 16.7%, while mean and max F-measure improve by 3.3% and 2.8%, respectively.
- Quantitative evaluation: On 360-SSOD, TDFNet ranks best on most metrics but second on max E-measure with a marginal gap from the top result.The best-performing metrics include S-measure, MAE, mean F-measure, max F-measure, and mean E-measure.
- Quantitative evaluation: On ODI-SOD, TDFNet achieves optimal performance across all metrics, including S-measure 0.876, MAE 0.028, mean E-measure 0.917, and mean F-measure 0.923.Compared with the second-best method, MAE is reduced by 20.0%, while mean E-measure and mean F-measure improve by 3.27% and 5.58%.
- Qualitative evaluation: Qualitative results show that TDFNet preserves clearer boundaries, suppresses non-target responses, and produces more coherent saliency maps for complex, low-contrast, and small-object scenes.The complete tri-branch model reduces false detections and boundary degradation observed in ERP-only and ERP-CMP configurations.
- Ablation study: LGF supports stable projection-adaptive ERP-CMP balancing through parameter-free latitude-based confidence masks, while its Mixed strategy adds 0.084M parameters and 0.086G FLOPs.The Geometry Only strategy is described as suitable for limited training data, whereas Mixed provides greater flexibility on larger datasets.
5 CONCLUSION
TDFNet addresses panoramic salient object detection by combining complementary ERP, CMP, and Tangent Projection representations. CDA aligns ERP and CMP through geometry-aware deformable attention, while LGF uses latitude-guided fusion and tangent references to refine cross-projection features.
- Conclusion: TDFNet introduces tri-projection deformable fusion for PSOD using ERP, CMP, and Tangent Projection representations.The design targets complementary global, local, and distortion-free semantic information.
- Conclusion: CDA establishes ERP-CMP spatial mappings and embeds them into deformable attention sampling to aggregate cross-projection context under projection deformation.This improves encoder robustness to projection deformation.
- Conclusion: LGF fuses ERP and CMP features and adaptively refines them with distortion-free Tangent Projection references to bridge local and global representations.Tangent Projection supplies local viewport information and is introduced into PSOD for the first time.