• Multi-oriental melon pose was rapidly estimated by low-cost weak localization. • Fruit-based localization can provide sufficient feature for pose estimation. • Achieved generalizability on domain-specific data by using low parameter model. • Fruit-pedicel pairs can be estimated in real time at 21.5 FPS on Orin Nano. Fruit pose is an important trait for robotic harvesting that provides not only control information for the robot arm but also indicates whether the fruit can be harvested. A pair of fruit and pedicel has been frequently used as the target object for estimating the fruit pose, and the pose estimation of multiple fruit-pedicel pairs within an image is also being focused on in current research to determine the targets to be harvested efficiently. The multi-fruit pose estimation is performed based on accurate object detection that requires a large amount of training data for feature representation of the object, especially pedicel, in complex background; however, there are spatio-temporal constraints in collecting various fruit-pedicel images, and precise labelling is a laborious task. Easier feature representation for localization can reduce model complexity and data-scale dependency. In this study, multi-fruit pose estimation of oriental melon was proposed, which includes a weakly fruit localization that reduces labelling effort while providing the fruit region containing the essential information for estimation of fruit-pedicel pose. The framework consists of a weak supervision-based localization and pose estimation models connected in sequence. The fruit regions are simultaneously and approximately localized, and each of these regions is then fed sequentially to the pose estimation. The proposed method has the advantage of being able to use a smaller model than existing models, which allows for achieving performance even with a small data scale and also improving real-time performance. The results showed that, compared to YOLO-pose models, our method achieved up to 0.23 higher PDK scores and exhibited an inference speed improvement of up to 8 fps. Our method minimized the effort required for localization and improved speed while demonstrating performance comparable to existing studies.