Albert Mundu

Scene Graph Generation has gained much attention in computer vision research with the growing demand in image understanding projects like visual question answering, image captioning, self-driving cars, crowd behavior analysis, activity recognition, and more. Scene graph, a visually grounded graphical structure of an image, immensely helps to simplify the image understanding tasks. In this work, we introduced a post-processing algorithm called Geometric Context to understand the visual scenes better geometrically. We use this post-processing algorithm to add and refine the geometric relationships between object pairs to a prior model. We exploit this context by calculating the direction and distance between object pairs. We use Knowledge Embedded Routing Network (KERN) as our baseline model, extend the work with our algorithm, and show comparable results on the recent state-of-the-art algorithms.

Image captioning is a challenging task in computer vision that automatically generates a textual description of an image by integrating visual and linguistic information, as the generated captions must accurately describe the image’s content while also adhering to the conventions of natural language. We adopt the encoder-decoder framework employed by various CNN-RNN-based models for image captioning in the past few years. Recently, we observed that the CNN-Transformer-based models have achieved great success and surpassed traditional CNN-RNN-based models in the area. Many researchers have concentrated on Transformers, exploring and uncovering its vast possibilities. Unlike conventional CNN-RNN-based models in image captioning, transformer-based models have achieved notable success and offer the benefit of handling longer input sequences more efficiently. However, they are resource-intensive to train and deploy, particularly for large-scale tasks or for tasks that require real-time processing. In this work, we introduce a lightweight and efficient transformer-based model called the Efficient Transformer Captioner (ETransCap), which consumes fewer computation resources to generate captions. Our model operates in linear complexity and has been trained and tested on MS-COCO dataset. Comparisons with existing state-of-the-art models show that ETransCap achieves promising results. Our results support the potential of ETransCap as a good approach for image captioning tasks in real-time applications. Code for this project will be available at https://github.com/albertmundu/etranscap.