Our SLURPP underwater restoration method demonstrates effective performance on clear image and medium predictions in using a single image as input. Our method is not designed for video underwater restoration, and we do not model the temporal consistency between underwater video frames. Nonetheless, we test our method on the MVK underwater dataset, which contains a diverse set of underwater videos under different aquatic conditions.
While reasonable view consistency and good restoration quality are observed in the foreground objects, in deep-sea environments flickering can occur in background regions. We note that on a single image basis, the restoration quality is good across depth ranges with no apparent artifacts. We believe the flickering stems from our method not modeling the temporal consistency between underwater video frames. Enhancing the model's temporal consistency for video applications remains a key direction for future research.
We show the results of our method on a sea bed scene with small depth variation as well as a scene in shallow water. Our method can produce view-consistent predictions of the sea snake, fish, and reefs in the videos, as well as the background seabed with good quality. Even with the movement of the sea snake, our method can produce consistent color restorations of its body. We note that we do observe flickering in the backscattering and transmission predictions. However, we do not see much flickering in the clear image predictions. We hypothesize that the reason for this is the scene lighting also flickers between frames, due to the complicated underwater environment and external lighting, such as the flash in the second fish video.
We show the results of our method on a deep sea video. Our method can produce view-consistent predictions of the fish in the foreground. However, we do observe some flickering in the low-light background, as well as in the backscattering and transmission predictions. Interestingly, at the start of the video, the fish and reefs are not in focus, and our transmission and backscattering predictions are also blurry. Once the fish and reefs are in focus, we see clear structures of the reefs and fish in the transmission and backscattering predictions.
We show the results of our method on a wreck scene. Our method can produce view-consistent predictions of the wreck in the video. However, we do observe strong flickering in all predictions, once the background is in view of the camera. We suspect that this effect is a combination of the lack of temporal modeling of our method, the floating dirt in the background negatively affecting implicit depth prior to the water medium prediction, and the low-light and noisy capture conditions of the video.