Image Compression Using Novel View Synthesis Priors

Real-time visual feedback is essential for tetherless control of remotely operated vehicles, particularly during inspection and manipulation tasks. Though acoustic communication is the preferred choice for medium-range communication underwater, its limited bandwidth renders it impractical to transmit images or videos in realtime. To address this, we propose a model-based image compression technique that leverages prior mission information. Our approach employs trained machine-learning based novel view synthesis models to store and render views of scenes that serve as priors, and minimizes the differences between camera images and model-rendered images to improve compressibility. We evaluate our method using data sets collected in an artificial ocean basin and demonstrate significant improvements in compression ratio and image quality compared to existing methods. Moreover, our approach exhibits robustness to the introduction of novel objects within the basin scene and maintains performance under real-world degradations, such as backscatter, as demonstrated on an online coral reef data set. These results highlight the promise of our proposed technique for enabling real-time, high-quality visual feedback over bandwidth-constrained underwater acoustic links.