Segmentation-Based Masked Sampling for text-to-animated image synthesis in disaster scenarios
Ru-Bin Won, Ji Hoon Choi, Minji Choi, Byungjun Bae · 2023
Generative AI has demonstrated significant capabilities in text-to-video synthesis, using advanced models characterized by extensive parameters. Given that current disaster alert services are mostly text-based and can be less accessible to many, this paper proposes a new approach. We aim to provide animated disaster images that precisely mirrors text prompts, thereby enhancing the efficiency and accessibility of disaster alerts. Our methodology combines the strengths of segmentation models and pre-trained Vision Transformer (ViT) mechanisms. By using a unique image selection based on CLIPScore and processing it with the CLIPSeg segmentation model, we generate an animated representation of disaster scenarios. This offers a simple, fast, and effective solution to the challenges of the current disaster alert systems.