A realistic protocol for evaluating weakly supervised object localization

In the field of deep learning, weakly supervised object localization (WSOL) represents a fascinating challenge: training models that learn to classify and locate objects within an image using only global class labels, without the need for costly bounding box (bbox) annotations. This approach is especially attractive for real-world applications, where obtaining thousands of manual bounding boxes is unfeasible. However, the evaluation of these models has historically relied on validation and test sets with annotated bboxes, which contradicts the premise of weak supervision. Recent research proposes a more realistic protocol that uses pseudo-boxes automatically generated through methods such as Selective Search, CLIP, or RPN, both for model selection and for estimating the binarization threshold of activation maps. This approach eliminates the need for human annotations during evaluation and achieves performance comparable to that obtained with real boxes. For companies seeking to implement computer vision solutions without relying on large volumes of labeled data, this line of research opens new possibilities. At Q2BSTUDIO we understand that transferring these advances to production environments requires robust and customized infrastructure. That is why we offer AI for businesses that integrates state-of-the-art models with workflows tailored to each sector. Our team develops custom applications capable of leveraging weak supervision techniques to reduce annotation costs and accelerate the deployment of visual inspection systems, assisted medical diagnosis, or automated image analysis. Furthermore, we combine these capabilities with AWS and Azure cloud services to ensure scalability, and with business intelligence services such as Power BI to visualize model performance. In a context where cybersecurity is critical, we also secure solutions through pentesting audits. Artificial intelligence applied to weak localization is not only an academic frontier but a practical tool that, with the right technical support, can transform industrial and business processes.