Enhancing Weakly Supervised Yield Estimation Through Learn-to-Pay-Attention Module

In recent years, the agricultural sector has witnessed the growing benefits of technological advancements, with computer vision emerging as a valuable tool to support cultivation activities. Specifically, computer vision has demonstrated its effectiveness in tasks such as counting and locating fruit, further enhancing yield estimation practices. As a result of limited fruit-specific datasets and the considerable resources required for their collection and annotation, weakly supervised methodologies have gained popularity. Recent studies demonstrated that these approaches can deliver outcomes comparable to those obtained with supervised approaches. Following these considerations, we developed a weakly supervised framework for counting and locating fruits on trees whose training relies solely on image-level presence/absence binary annotations. Our network incorporates a Learn-To-Pay-Attention module, optimizing the network and generating binary predictions used in the loss computation. The experimental results, conducted on images of mango cultivation, showcase significant enhancements in terms of robustness and accuracy compared to State-of-the-Art approaches.