An arXiv preprint reports that MIMONet posted better evaluation scores than competing methods in most of its salient-object detection comparisons, with near-best performance in the remaining cases. The study targets situations in which target sizes vary. The system creates three input resolutions and produces one saliency map per branch. These are point-estimate comparisons; no confidence intervals or statistical significance tests are reported.
The document is marked arXiv:2608.25733v1 [cs.CV] and dated 26 Aug 2026.
A three-scale design
MIMONet creates three versions of each input: the original, one scaled to 0.5 of its size and another scaled to 1.5. The three streams pass through a shared-parameter ResNet50 backbone, exchange information between scales and produce one saliency map per branch.
Inside the network, its multi-scale perception module, or MSP, splits each feature layer into three sub-layers in a 1:2:1 ratio. It applies 3 × 3, 5 × 5 and 7 × 7 convolutions, uses scale transformations and aggregates the results.
Each saliency map is trained with binary cross-entropy and intersection-over-union losses. A joint saliency loss, or JSL, supplies common supervision across the multiple outputs and boundary supervision.
Where it was tested
Experiments used six named image datasets: ECSSD with 1,000 images, PASCAL-S with 850, HKU-IS with 4,447, DUT-OMRON with 5,168, DUTS with 10,533 training images and 5,019 test images, and the SOC test set with 1,200 images.
Training used the DUTS training set. Inputs were fixed at 352 × 352 pixels, with random rotations and horizontal flips used as augmentation. The model trained in batches of 32 with Adam, an initial learning rate of 0.00004, cosine annealing and 50 epochs.
For its main comparison, the study measured MIMONet against 21 recent state-of-the-art methods. Although the model produces outputs at several scales, the final evaluation used only the original-size saliency maps.
The reported scorecard
The paper reports six saliency measures: maximum, average and weighted F-measures; S-measure; E-measure; and pixel-wise mean absolute error. In the first metric block for MIMONet, the values are 0.956, 0.941, 0.932, 0.933, 0.962 and 0.027, respectively. The first five measures favor higher values, while the error measure favors a lower value.
The paper also reports 28.36 million parameters and an inference speed of 32 frames per second. The authors conclude that MIMONet reaches state-of-the-art performance on multiple datasets with relatively few parameters.
What changed inside the model
In the input ablation, the reported three-input configuration—low, original and high resolution—had higher accuracy than the two-input version. The five-input version was worse than the three-input version in some cases.
Tests of MSP favored the proposed configuration in the reported comparisons. Among the split ratios tested, 1:2:1 had the strongest reported performance; removing scale transformations worsened detection, and MSP had better reported results than the alternative PPM and ASPP modules.
The setting that included JSL had better reported detection performance than the setting without it. For the boundary construction, the study selected four branches for the maximum reported benefit.
The authors’ next step
The authors identify point-cloud saliency detection as a direction for future work.
The abstract states that the model code will be released. The work reports support from the Natural Science Foundation of China, Shenzhen programs and the CCF-Lenovo Blue Ocean Research Fund.
Paper data and sources
Original title: MIMONet: Multi-scale Input and Multi-scale Output Network for Salient Object Detection
Authors: Zhaojian Yao, Wei Gao, Tiesong Zhao et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text