Advertisement

ShipSense
ShipSense
New method for augmenting synthetic image data in a few shots using stable diffusion fine-tuned for any object. ShipSense used custom stable streaming to create realistic synthetic image data of ships and trained convolutional neural networks (CNNs) to detect and localize ships from satellite images. They also built a data visualization platform for stakeholders to monitor overfishing. To improve this platform, they identified several hotspots of suspicious dark vessel activity by searching through more than 55,000 AIS radar recordings.
While people have previously tried to create AI models to detect overfishing, the accuracy was poor due to high class imbalance. There are few positive examples of ships on the water, compared to the countless negative examples of bodies of water without ships. Researchers have used GANs to generate synthetic data for other purposes. However, it takes around 50,000 image samples to train a decent GAN. The largest satellite dataset contains only about 2,000 samples.
They realized that Stable Diffusion (SD), a popular text-image AI model, could be reused to generate an unlimited number of synthetic ship image data based on relatively few inputs. They were able to obtain very realistic synthetic images using only 68 original images.
Vote :










