TDFNet: Tri-projection Deformable Fusion Network for Panoramic Salient Object Detection

Qiangqiang Zhou; Jiacong Yu; Jiawei Xu; Yong Chen; Xin Huang; Ping Li

cs.CVAug 26, 2026

TDFNet: Tri-projection Deformable Fusion Network for Panoramic Salient Object Detection

Authors: Qiangqiang Zhou, Jiacong Yu, Jiawei Xu, Yong Chen, Xin Huang, Ping Li

Organizations: School of Artificial Intelligence, Jiangxi Normal University, Nanchang, China

Abstract

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.

Figures & tables

Figure 1. Comparison of projection representations used in PSOD. (a) ERP maintains complete spherical topology but suffers from polar stretching distortion. (b) CMP reduces polar deformation through hexahedral decomposition but introduces semantic discontinuities at cube-face boundaries. (c) Tangent Projection, introduced in this work, provides distortion-reduced local perspective views through discrete viewport mapping, complementing ERP and CMP with high-fidelity geometric details. projection

Figure 2. Overall architecture of the proposed TDFNet. The network incorporates three core components: (i) a multi-view tangent projection branch for distortion-free local feature extraction, (ii) the cross-projection deformable attention (CDA) deployed in both ERP and cubemap branches to enhance robustness against projection distortions, and (iii) the latitude-guided fusion (LGF) module that progressively integrates tri-branch features under spherical geometric guidance for cross-projection alignment. our method

Figure 3. The structure of the deformable Transformer block in the CDA module. Deformable Transformer block in CDA.

Figure 4. The structure of the LGF module. LGF module

Model	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$maxE\uparrow$	$meF\uparrow$	$maxF\uparrow$
FA ( Huang et al., 2020 )	.826	.021	.885	.900	.748	.770
DDS ( Li et al., 2019 )	.799	.023	.865	.904	.695	.722
DT ( Zhao et al., 2023 )	.849	.017	.907	.919	.774	.793
GFI ( Zhu et al., 2021 )	.831	.021	.894	.900	.753	.769
SCW ( Yu et al., 2021 )	.830	.021	.886	.887	.754	.759
BIPG ( Yao and Wang, 2021 )	.811	.024	.885	.890	.727	.740

Table 1. Quantitative comparison results on 360-SOD dataset. “Ours” adopts the Geometry Only strategy in the LGF module.

\uparrow

and

\downarrow

indicate that the larger scores and the smaller ones are better, respectively. The best scores are highlighted in bold .

Model	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$maxE\uparrow$	$meF\uparrow$	$maxF\uparrow$
LD ( Huang et al., 2023 )	.756	.034	.845	.863	.492	.511
FA ( Huang et al., 2020 )	.717	.039	.727	.735	.520	.532
HVP ( Liu et al., 2020 )	.773	.143	.800	.854	.500	.529
DT ( Zhao et al., 2023 )	.770	.026	.832	.865	.644	.657
GFI ( Zhu et al., 2021 )	.767	.034	.829	.848	.526	.536
SCW ( Yu et al., 2021 )	.760	.029	.813	.854	.511	.518

Table 2. Quantitative comparison results on 360-SSOD dataset. “Ours” adopts the Mixed strategy in the LGF module.

\uparrow

and

\downarrow

indicate that the larger scores and the smaller ones are better, respectively. The best scores are highlighted in bold .

Model	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$maxE\uparrow$	$meF\uparrow$	$maxF\uparrow$
View ( Wu et al., 2022 )	.831	.035	—	—	—	.822
FA ( Huang et al., 2020 )	.730	.050	.771	.790	.610	.632
DDS ( Li et al., 2019 )	.791	.045	—	—	—	.761
HVP ( Liu et al., 2020 )	.732	.061	.787	.795	.616	.627
DC ( Zhu et al., 2023 )	.659	.120	.685	.691	.520	.530
MDSAM ( Gao et al., 2024 )	.706	.084	.744	.748	.597	.603

Table 3. Quantitative comparison results on ODI-SOD dataset. “Ours” adopts the Mixed strategy in the LGF module.

\uparrow

and

\downarrow

indicate that the larger scores and the smaller ones are better, respectively. The best scores are highlighted in bold .

Model	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$meF\uparrow$
FA ( Huang et al., 2020 )	.621	.056	.700	.345
DDS ( Li et al., 2019 )	.612	.057	.700	.325
ACCo ( Li et al., 2022 )	.613	.060	.694	.334
MPFR ( Cong et al., 2023 )	.617	.052	.742	.370
DT ( Zhao et al., 2023 )	.611	.058	.652	.310
CPNet ( Wen et al., 2025 )	.651	.051	.729	.382

Table 4. Quantitative comparison results on F-360iSOD dataset. “Ours” adopts the Geometry Only strategy in LGF.

\uparrow

and

\downarrow

indicate that the larger scores and the smaller ones are better, respectively. The best scores are highlighted in bold .

Figure 5. Qualitative comparisons of various PSOD methods under several challenging scenarios. The all-black images marked with yellow boxes indicate that the corresponding saliency maps cannot be obtained. Qualitative comparisons

360-SOD	Settings	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$meF\uparrow$	360-SSOD	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$meF\uparrow$
	ERP	.824	.026	.889	.742		.748	.039	.827	.595
	ERP-CMP	.858	.018	.918	.812		.793	.028	.862	.677
	ERP-CMP-TAN	.878	.016	.925	.822		.796	.027	.872	.689

Table 5. Ablation study on different branch combinations in the proposed model.

Figure 6. Visualization of three variants, ERP, ERP-CMP and ERP-CMP-TAN. Visualization of three variants

Figure 7. Visualization of multi-projection features extracted by the tri-branch encoder. Visualization of different modules.

Figure 8. Visualization results of ablation studies in different modules. Visualization of different modules.

360-SOD	Settings	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$meF\uparrow$	360-SSOD	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$meF\uparrow$
	w/o CDA	.869	.017	.923	.807		.795	.027	.865	.677
	w/o LGF	.878	.016	.925	.822		.796	.027	.872	.689
	Ours	.883	.015	.927	.835		.804	.026	.877	.697

Table 6. Ablation experiment of each module in the model.

Strategy	360-SOD				360-SSOD				Complexity
Strategy	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$meF\uparrow$	$S_{m}\uparrow$	$MAE\downarrow$	$meE\uparrow$	$meF\uparrow$	Params (M)	FLOPs (G)
Geometry Only	0.883	0.015	0.927	0.835	0.789	0.029	0.863	0.678	0.000	$\sim$ 0
Mixed	0.860	0.019	0.917	0.802	0.804	0.026	0.877	0.697	0.084	0.086

Table 7. Ablation study on the fusion strategies of the LGF module.

Figure 9. Visualization of feature enhancement by the proposed CDA module.

Figure 10. Visualization of the LGF fusion process and final predictions.

Salient Object Detection

Salient Object Detection finds visually prominent objects in images or video and predicts their pixel-level foreground masks.

See all 37 papers

SphereSOD: Geometry-Structure Coupled Learning for 360 Salient Object Detection
Sep 7, 2026Junsong Zhang, Zhijie Shen, Shuai Zheng +4Salient Object Detection
DDS
Learning Projection-Aware 360-Degree Image Rectification via Dual-Projection Fusion
Nov 30, 2025Yuhao Shan, Qianyi Yuan, Jingguo Liu +3Parameter Estimation
college of electronic and information engineering, Southwest University, Chongqing 400700, China
Spherical Geometry-Aware Panoptic Scene Graph Generation for Equirectangular Panoramas
Oct 20, 2026Zhe Wu, Xiaoyan Fei, Xinyi Jiang +4Scene Graph Generation Panoptic Segmentation
Shanghai DianJi University, Shanghai, China · University of Minnesota Twin Cities, Minneapolis, MN, USA

Sep 7, 2026J/K move · Enter open · S save

SphereSOD: Geometry-Structure Coupled Learning for 360 Salient Object Detection

Junsong Zhang, Zhijie Shen, Shuai Zheng, Feng Li+3

DDS

Salient Object Detection

Open paper AskPDF arXiv

360° salient object detection (SOD) aims to accurately segment salient regions across a full field of view. However, equirectangular projection (ERP) introduces severe spatial distortion when mapping the spherical domain onto a planar representation. Existing methods mainly focus on compensating projection distortion while overlooking the interaction between panoramic geometry and salient object structure during feature perception and prediction refinement. To this end, we propose SphereSOD, an ERP-native framework that couples panoramic geometry with evolving salient structures. Specifically, spherical geometry governs feature sampling and spatial weighting, while coarse-grained saliency and contour prediction influence context aggregation during the progressive decoding process. SphereSOD first initializes deformable sampling based on spherical projection geometry and then employs bounded, content-adaptive offsets, yielding features that are better aligned with the underlying panoramic geometry. Subsequently, the decoder performs structure-guided context aggregation and progressive refinement to recover complete salient regions and accurate boundaries. Extensive experiments on three public 360° SOD benchmarks demonstrate state-of-the-art performance and a favorable accuracy-efficiency trade-off, supporting structurepreserving inference directly in ERP space as a promising alternative to projection-heavy panoramic pipelines.

Abstract

Salient Object Detection

Explore similar work

SphereSOD: Geometry-Structure Coupled Learning for 360 Salient Object Detection

Learning Projection-Aware 360-Degree Image Rectification via Dual-Projection Fusion

Spherical Geometry-Aware Panoptic Scene Graph Generation for Equirectangular Panoramas