ETD

• ai
• computer vision
• generative ai
• llm
• neural network
• scene understanding

Scene understanding is a fundamental task in computer vision and recent advancements in this field have been driven by the development of Multimodal Large Language Models (MLLMs). However, these models can suffer from hallucinations and are not entirely reliable, especially when it comes to a sequence of images, such as video data. This work proposes a novel approach that combines the strengths of MLLMs with an Object Detection model and a graph-based representation of the scene to enhance the reliability and coherence of the outputs generated by MLLMs.