A preprint reports that TDFNet, a model using equirectangular (ERP), cubemap (CMP) and tangent-projection branches, delivered strong reported results across four panoramic salient-object-detection benchmarks. On 360-SOD, the authors list an S-measure of 0.883, mean absolute error (MAE) of 0.015, mean E-measure of 0.927, maximum E-measure of 0.932, mean F-measure of 0.835 and maximum F-measure of 0.847. In the comparison with SCFA, TDFNet's reported MAE was 16.7% lower, while its mean and maximum F-measures were 3.3% and 2.8% higher.
The design asks whether combining the three projections through deformable fusion can address distortions in panoramic salient-object detection. The document is identified as arXiv:2608.25808v1 and dated 26 August 2026.
Three representations, one model
TDFNet has three branches: ERP, cubemap and tangent projection. Its cross-projection deformable attention, or CDA, is used to enhance ERP and cubemap features, while latitude-guided fusion, or LGF, progressively integrates information from all three representations.
The research question focuses on two distortions named in the analysis: polar stretching in ERP and discontinuities at cubemap face boundaries. The paper examines the proposed fusion design through benchmark comparisons and ablations.
What the benchmarks showed
The evaluation covered four datasets: 360-SOD, with 500 images; 360-SSOD, with 1,105 images in 10 semantic categories; F-360iSOD; and ODI-SOD. F-360iSOD is described as an instance-level set with 107 images, 1,165 salient objects and 72 categories. ODI-SOD contains 6,263 images with a minimum resolution of 2K. The study followed the original train-test splits for all four datasets.
For comparison, the authors included 19 methods: nine designed for panoramic images and 10 for ordinary 2D images. Training used ERP inputs resized to 640 by 1,280 pixels, with no data augmentation, a batch size of two and 50 epochs.
Results on 360-SSOD were strong but not first across every measure. The Mixed-strategy row reported an S-measure of 0.804, MAE of 0.025, mean E-measure of 0.877, maximum E-measure of 0.883, mean F-measure of 0.697 and maximum F-measure of 0.708. The paper says the model ranked second on maximum E-measure, with only a marginal gap from the top result.
On ODI-SOD, the authors report optimal performance across the listed metrics: S-measure 0.876, MAE 0.028, mean E-measure 0.917, maximum E-measure 0.923, mean F-measure 0.833 and maximum F-measure 0.845.
On F-360iSOD, the Geometry Only configuration was reported as best across the listed metrics, with an S-measure of 0.745, MAE of 0.030, mean E-measure of 0.762 and mean F-measure of 0.545. Compared with the second-best method, the authors report a 20.0% MAE reduction, a 3.27% mean E-measure improvement and a 5.58% mean F-measure improvement.
The full setup scored highest in ablations
In the reported branch-ablation tables for both datasets, the complete ERP-CMP-TAN configuration had higher scores than the ERP-only and ERP-CMP variants across the listed metrics.
The module comparisons showed lower reported metrics for configurations without CDA or LGF, with the larger degradation appearing in the CDA-removal configuration. CDA is the cross-projection deformable-attention component, while LGF is the component that progressively integrates the three representations.
The strategy varied by dataset
The LGF strategy comparison reported Geometry Only values of 0.883 for S-measure, 0.015 for MAE, 0.927 for mean E-measure and 0.835 for mean F-measure on 360-SOD. Its Mixed entry for 360-SSOD reported 0.804, 0.026, 0.877 and 0.697 on those measures, along with 0.084 million parameters and 0.086 billion FLOPs. That 0.026 MAE comes from the separate LGF-strategy comparison; the 360-SSOD benchmark row reports 0.025 for its Mixed strategy.
In visual comparisons, the authors describe clearer boundaries, fewer non-target responses and coherent saliency maps for scenes containing multiple, low-contrast and small objects.
The caveats behind the scores
The benchmark figures are descriptive comparisons. The supplied analysis reports no confidence intervals, significance tests, repeated-run variability or other uncertainty estimates, so the reliability of the reported differences across repeated runs is not quantified.
Evaluation was limited to the four named benchmarks and their original splits. The model used ERP-format inputs and no data augmentation; the paper does not describe a common retraining protocol for competing methods, and competitor maps were author-provided or generated with author code.
Because different LGF strategies were used across datasets, the cross-dataset values do not represent one fixed configuration. The analysis identifies fixed-seed replication with independently retrained baselines as a needed next test of reproducibility, while the authors state that their code will be made publicly available.
A benchmark study, not a deployment result
As a preprint, the work offers an architectural proposal and benchmark evidence for researchers developing panoramic salient-object-detection systems. The reported evaluation does not establish performance in robotic or virtual-reality tasks, generalization beyond the four tested benchmarks, results under non-ERP input protocols or an effect from data augmentation.
Paper data and sources
Original title: TDFNet: Tri-projection Deformable Fusion Network for Panoramic Salient Object Detection
Authors: Qiangqiang Zhou, Jiacong Yu, Jiawei Xu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text