MoE-RelationNet++: Adaptive Keypoint Selection via Conditional Experts
Year: 2026
Type: article
Abstract: Modeling contextual relationships among key features is crucial for improving object detection, yet existing relation-based methods rely on fixed feature selection and shared transformations, limiting their ability to capture diverse feature interactions in complex scenes. To address this, we propose MoE-RelationNet++, a relation enhancement framework based on a Mixture-of-Experts (MoE) mechanism. Unlike fixed selection and shared transformations, the proposed MoE enhancement module adopts a conditional computation paradigm: a dynamic router adaptively assigns features to specialized experts, enabling heterogeneous relationship modeling and overcoming the representational bottlenecks inherent in traditional shared mappings. Furthermore, to alleviate its computational burden and reduce redundant inputs, a lightweight key selector using depthwise separable convolution is introduced to adaptively identify informative features. To ensure robust relation modeling and prevent noisy or unreliable feature interactions from degrading the experts, an energy verification mechanism is employed to evaluate feature reliability and refine the overall process. Extensive experiments on MS COCO show consistent improvements across multiple detectors, increasing AP by 4.2, 3.2, 3.3, and 3.1 points for RetinaNet, FCOS, ATSS, and Faster R-CNN, respectively. Additionally, the method achieves a 1.5 AP gain on the VisDrone-DET2019 benchmark. These results demonstrate that MoE-RelationNet++ effectively captures heterogeneous relations via conditional expert routing, overcoming the representational limitations of fixed transformations. Moreover, it can be seamlessly integrated into various detection frameworks as an add-on enhancement module, consistently improving their performance without modifying the base architecture.
Source: Applied Sciences
DOI: 10.3390/app16115192
Authors: Yuhan Peng, Gaofeng Zhang
Institution: Lanzhou University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
Funder: Lanzhou University
Award: G7907765824
OpenAlex ID: W7162074424