In agricultural harvesting robots, fruit pose information is essential for determining the grasping point. However, in greenhouse environments, frequent occlusion caused by leaves and stems considerably degrades the accuracy of existing models. This study aims to improve cucumber pose estimation performance in occluded scenes. To achieve robust estimation performance under various occlusion conditions, we designed an approach that infers the relative offsets of occluded keypoints from visible keypoints and reconstructs the entire pose. The training dataset consisted of images collected from an actual cucumber greenhouse under diverse capturing angles, while the test dataset was constructed by imposing various occlusion conditions using leaf patches to emulate realistic field distribution. In addition, a similarity-based offset loss was introduced to encourage consistent relative-offset regression among keypoints. For practical deployment in real-world applications, a YOLO (you only look once)-based model was adopted to simultaneously detect cucumbers, estimate keypoints, and regress the relative offsets between keypoints within a single architecture. Comparative experiments demonstrated that the proposed method achieved higher accuracy than existing approaches across various occlusion scenarios. These results indicate that the approach is capable of compensating for missing pose components by reasoning from partially visible regions. With continued research, this method is expected to enhance the precision of vision pipelines for cucumber harvesting robots operating under real greenhouse conditions.