The idea behind this project is a good one: murky water wrecks footage, and the maker spent six months of solo effort building Visual Recovery, a program that cleans up underwater video live. The algorithm was designed specifically for real-time use in murky conditions, and the whole thing is still experimental.
Technically, everything runs through NVIDIA CUDA on the GPU and targets 4K at 60 fps. What stands out to me is the hardware choice: it all runs on a laptop with an integrated graphics controller rather than a desktop card. The footage comes from a camera on a remotely operated underwater vehicle that links to its controller through a tether, with the controller supplying the Wi-Fi. The video bitrate was also pushed up to 90 Mbit/s from the factory default of 60.
The numbers are worth noting. Processing takes roughly 3 to 10 milliseconds per frame, with about 5 to 10 ms of GPU time in the tests shown. An adjustable safety buffer, never set below 90 ms, absorbs timing jitter, and total latency landed near 100 ms during testing. Quick processing keeps frame-to-frame color consistent, nothing stuttered during the tests at the water's edge, and the laptop held a steady 4K 60 fps even off the charger. In the comparison, a divider line wipes across the frame between the before and after views while the ROV slowly advances and turns above the seabed; the enhanced side shows better color balance, more contrast, and clearer rust, sediment and structural detail.
A couple of open points remain. Total latency has not yet been precisely measured, and with no access to the drone manufacturer's SDK, the feed depends on the drone's Wi-Fi. Sunlight direction and intensity were the biggest challenge in testing. The maker is keen to hear from ROV operators and anyone doing underwater machine vision work.