CSSGNet: Co-salient space guidance based visual servoing network

Background interference limits the generalization of learning-based relative-pose estimation for visual servoing. This paper proposes CSSGNet, a co-saliency space-guided visual servoing network. A group-conditioned Co-Saliency Feature Mapper jointly processes the current and desired images as a composite input. Global channel aggregation and spatial attention suppress background interference while preserving foreground positions and contours. The resulting foreground maps are processed by a Siamese Vision Transformer to estimate the relative pose, which drives a classical position-based visual servoing controller. In an offline evaluation of 100 image pairs acquired in an unknown real-world environment, CSSGNet achieves mean axis-wise translation estimation errors of 0.8746–0.9015 mm and an SO(3) geodesic rotation estimation error of 0.0138 ± 0.0010 rad without real-world fine-tuning. These statistics describe prediction errors across image pairs, rather than residual robot-positioning errors after closed-loop execution. The image-pair evaluations support improved pose-estimation performance under the evaluated scene changes, including sim-to-real transfer. Separate representative closed-loop experiments illustrate the feasibility of integrating CSSGNet with classical PBVS control. The code is available at https://github.com/zhaoweishuo/CSSGNet . • A co-saliency space suppresses scene-specific background interference. • A Siamese Vision Transformer regresses relative pose from co-saliency maps. • Real-world tests achieve submillimeter mean axis-wise translation errors.