Multi-Class Label Detection and Bounding Box Regression Using Transformer with a Customized Loss Function

This paper introduces an innovative approach that utilizes Vision Transformers in conjunction with a loss function derived from Intersection over Union (IOU) and Mean Squared Error (MSE) for the training and prediction of class labels and their associated bounding boxes. Our experiments, conducted on a dataset containing two distinct class labels, demonstrate an impressive 96% accuracy in label prediction and a 95% IoU accuracy for bounding boxes.