Rethinking False Positives with Class Activation Maps: Can we define false positives without ground-truths?

In this study, we have used Class Activation Mapping (CAM) technique to improve the accuracy of our Gloves classification model. By extracting the CAM and computing its areas, we have created a mechanism that effectively reduces false positives. We have trained the ResNet18 model on 2000 images of hands with and without gloves, achieving 98.65% accuracy. We have used the last convolutional layer of the model to extract the CAM outputs using the Gradient-CAM technique. Then, we have taken the extracted areas into account to decide whether a hand exists in the given image. Our results show that this approach is effective in reducing false positive detections.