Beneath the surface of an aquaculture tank, two cameras watch the same fish from slightly different angles, and a new method reads their paired views to judge its length, width and weight. Researchers at Sultan Qaboos University in Oman and partner institutions have shown that such measurement can be done without the heavy labor of reconstructing the whole watery space around the fish. In a long study of tilapia across fifteen months of growth, the approach proved both quick and reasonably faithful, though its reach remains bounded: it has yet to tally every fish or account for an entire farm,
Stereo Cameras and AI Estimate Fish Weight in Real Time Without Full Depth Maps
Depth only where the fish are, not across the whole scene
Mark
Why skip the full depth map? Isn't that the standard way to do stereo vision?
Mimi
It is, but a dense depth map computes distance for every pixel in the scene. Most of that is water. StereoGeoBox only works out depth where a fish has been detected, which is why it needs so little computing.
Luke
And the memory figure shows it: 17.7 megabytes of GPU memory. Worth noting that's the researchers' own measurement on one platform, the NVIDIA Jetson Thor.
Mark
How well did it actually estimate size?
Mimi
In a controlled test, fish with a reference length of 15.30 centimeters averaged 15.25 in the system's estimate. Across most datasets it had the lowest or near-lowest errors, with fewer outliers.
Luke
Careful with that. 'Most datasets' and 'near-lowest' mean it didn't win everywhere. And the 15.25 figure comes from one controlled test, not the whole 15 months.
Mark
What about murky water or bubbles?
Mimi
It kept reasonable accuracy, but when degradation got severe, fewer measurements were usable. It didn't fail so much as thin out.
Luke
'Reasonable' isn't quantified in what we have. We don't know the error under those conditions.
Mark
So can a farm use this to know its total stock?
Mimi
Not yet. Frames are processed separately, so the same fish can be counted more than once, and the study didn't estimate whole-farm biomass.
Luke
Right, and it was only tilapia. Other species and commercial farms are untested.
O Pulso
- Feeding, growth and harvest decisions hinge on weight estimates that conventional underwater imaging struggles to deliver quickly amid reflections, bubbles and murky water.
- The team pushed a shortcut into motion: detect fish with YOLO11 in both camera views and compute depth only where the fish appear.
- At 25.40 frames per second on an NVIDIA Jetson Thor, the system keeps pace with real-time monitoring while holding steady under shifting light, bubbles and turbidity.
- Severe visual degradation still cuts usable measurements, and repeat sightings of the same fish remain unresolved.
- The method is landing as a promising prototype, with testing on other species and commercial farms still ahead before it can be judged in the field.
Beneath the surface of an aquaculture tank, two cameras watch the same fish from slightly different angles, and a new method reads their paired views to judge its length, width and weight. Researchers at Sultan Qaboos University in Oman and partner institutions have shown that such measurement can be done without the heavy labor of reconstructing the whole watery space around the fish. In a long study of tilapia across fifteen months of growth, the approach proved both quick and reasonably faithful, though its reach remains bounded: it has yet to tally every fish or account for an entire farm, and it awaits trial beyond the research setting.
Two synchronized cameras sit underwater, each looking at the same tilapia from a slightly different angle. From those paired images, a system called StereoGeoBox works out how long the fish is, how wide, and how much it weighs, without building a full three-dimensional map of the water around it.
The work comes from researchers at Sultan Qaboos University in Oman and collaborating institutions, published on October 23, 2026, in Computers and Electronics in Agriculture. For fish farmers, reliable weight estimates guide feeding, growth tracking and harvest planning, yet conventional underwater imaging demands heavy computing and falters amid reflections, bubbles and cloudy water.
StereoGeoBox takes a shortcut. An AI detector finds fish in both images, and depth is calculated from the geometry of the boxes around each fish. Length and width then feed a machine-learning model that predicts weight. Over fifteen months of monitoring tilapia at four growth stages, the system produced the lowest or near-lowest errors in most datasets, and in one controlled test estimated a 15.30-centimeter fish at 15.25 centimeters.
Running at 25.40 frames per second, it kept pace with real-time monitoring and held reasonably well under changing light, bubbles and turbidity, though severe degradation reduced usable measurements.
The authors are candid about the limits. The study did not count every fish or estimate whole-farm biomass, and the same fish may be seen more than once. Testing on other species and in commercial aquaculture is still required.
Two synchronized cameras sit underwater, each looking at the same tilapia from a slightly different angle. From those paired images, a system called StereoGeoBox works out how long the fish is, how wide, and how much it weighs, without ever building a full three-dimensional map of the water around it.
The work comes from researchers at Sultan Qaboos University in Oman and collaborating institutions, and it was published online on Oct. 23, 2026, in Computers and Electronics in Agriculture. The paper is titled "Real-time in-water fish biomass estimation from stereo bounding-box geometry without dense depth maps." The research involved the university's Department of Electrical and Computer Engineering, Abdullah Al Salem University in Kuwait, and Oman's Ministry of Agriculture, Fisheries and Water Resources.
The problem it addresses is a practical one for fish farmers. Reliable weight estimates help operators manage feeding, track growth and plan harvests. But conventional underwater imaging tends to demand a lot of computing power, and it can struggle with reflections, bubbles and cloudy water.
StereoGeoBox takes a shortcut. It uses YOLO11, an AI detection model, to find fish in both images of a pair. It then calculates depth from the geometric relationship between the bounding boxes drawn around the same fish in each view. Estimated length and width go into an XGBoost machine-learning model, which predicts weight. In the words of corresponding author Dr. Said Al-Abri, estimating depth only where fish are detected cuts computational demands while keeping the speed and accuracy that real-time monitoring needs.
To test it, the team monitored tilapia over 15 months at four growth stages: 2, 3, 6 and 15 months. They used two stereo-camera systems, with manual measurements of length, width and weight serving as reference values. Across most datasets, StereoGeoBox produced the lowest or near-lowest errors compared with conventional stereo-vision and learning-based depth methods, and it generated fewer outliers. In a controlled test on fish with a reference length of 15.30 centimeters (6.02 inches), its average estimate was 15.25 centimeters (6.00 inches).
Speed was the other measure. Running with an FD camera on an NVIDIA Jetson Thor platform, the system processed 25.40 frames per second, with a latency of 28.2 milliseconds and GPU memory use of 17.7 megabytes. The prototype paired underwater cameras with edge computing and a cloud-based dashboard, so growth estimates could be recorded and viewed remotely.
The system held up reasonably well when lighting changed and when oxygen bubbles and turbidity were introduced. Severe visual degradation, though, reduced the number of usable measurements.
The study also has clear limits. It did not demonstrate automatic counting of every fish, nor independent estimation of a whole farm's total biomass. Because each frame was processed separately, the same fish could be observed more than once. The authors note that further testing on other species and in commercial aquaculture settings is still required before the approach can be judged outside the research setting.
Citações Notáveis
By estimating depth only where fish are detected, StereoGeoBox reduces computational demands while retaining the speed and accuracy needed for real-time aquaculture monitoring.— Dr. Said Al-Abri, corresponding author, Sultan Qaboos University
Quer a matéria completa?
Leia o original em Phys.org ↗
Technology