The paper introduces a method for generating unambiguous descriptions (referring expressions) of specific objects or regions in images and for interpreting such descriptions to infer the object being described. The authors propose a model that combines convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to handle real images and text, outperforming previous methods that do not consider potentially ambiguous objects in the scene. They also present a new large-scale dataset for referring expressions, based on the MS-COCO dataset, and release a toolbox for visualization and evaluation. The model is trained using a semi-supervised approach, where descriptions are automatically generated for image regions. The paper discusses the evaluation metrics used for the generation and comprehension tasks, including precision@1 and human evaluation, and compares different training methods. The results show that the proposed model outperforms baseline methods in both tasks, demonstrating the importance of considering the listener's perspective in generating unambiguous descriptions.The paper introduces a method for generating unambiguous descriptions (referring expressions) of specific objects or regions in images and for interpreting such descriptions to infer the object being described. The authors propose a model that combines convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to handle real images and text, outperforming previous methods that do not consider potentially ambiguous objects in the scene. They also present a new large-scale dataset for referring expressions, based on the MS-COCO dataset, and release a toolbox for visualization and evaluation. The model is trained using a semi-supervised approach, where descriptions are automatically generated for image regions. The paper discusses the evaluation metrics used for the generation and comprehension tasks, including precision@1 and human evaluation, and compares different training methods. The results show that the proposed model outperforms baseline methods in both tasks, demonstrating the importance of considering the listener's perspective in generating unambiguous descriptions.