FishSense Lite: A Camera-Based, Single Laser Hardware Framework with Software System for In Situ Fish Length Measurement
Authors: Christopher L. Crutchfield, Kyle Hu, Nathan Hui, Omair Qazi, Jennifer Loch, Alli Candelmo, Brice X. Semmens, Curt Schurgers, Ryan Kastner
Organizations: University of California, San Diego, La Jolla, California, USA · Reef Environmental Education Foundation, Key Largo, Florida, USA · Scripps Institution of Oceanography, La Jolla, California, USA
Fish stock assessments need fish lengths paired with species from far more places than trained surveyors can reach. Recreational divers already visit many of those places, but measuring a fish underwater has required trained observers or specialist equipment. We present FishSense Lite, which recovers fish length from a single photograph taken with a recreational dive camera, a dive laser pointer, and a 3D-printed mount. The diver aims the laser at a fish and takes a picture; afterward, the dot's position in the image gives the range to the fish, and the range turns the fish's size in the image into a length. A citizen-science program, trained by the system's builders, taught its own divers and collected lengths paired with species for 73 wild fish across seven open-water dives. On seven wild fish also measured by an underwater stereo-imaging system, the two instruments agreed without detectable bias (mean difference +0.8%). On rigid targets of known length, FishSense Lite recovered every target to within 1.3%, and 99% of 1,055 frames fell within a 15% error budget. From the error sources we measure, we derive a field protocol: calibrate against a dive slate on every dive, photograph fish within 15 degrees of perpendicular and beyond 1 m, and take at least ten frames of each.
Figures & tables
Figure 1. The FishSense Lite system adds a laser pointer to the Olympus TG-6 commercial-off-the-shelf (COTS) dive camera. A recreational diver takes a picture of a fish with the laser pointed at it. Our software calculates the fish length using only that information. A photo of the \fslSystem.
Technique
Barrier to a volunteer
Accuracy
Range
Stereo Imagery
two calibrated cameras; recalibrated whenever the rig shifts ( Goetze et al., 2019 ; Delacy et al., 2017 )
−0.6 cm mean error ( Harvey et al., 2001 )
2–10 m ( Mallet and Pelletier, 2014 )
Acoustic Methods
Specialist instrument ( Mueller et al., 2006 )
±0.5 cm at 1 m to ±6 cm at 12 m ( Mueller et al., 2006 )
1–12 m ( Mueller et al., 2006 )
Laser Caliper
Consumer components; lasers kept parallel ( Heppell et al., 2012 ; Stock et al., 2021 )
5–7% ( <10∘ off axis) ( Stock et al., 2021 )
0.5–5 m ( Heppell et al., 2012 ; Stock et al., 2021 )
FishSense Lite (this paper)
Consumer components and a printed mount; slate photographed each dive
−2.2% median per frame (pool); 8.9% vs. stereo (Section 5 )
0.5–5 m, beyond 1 m recommended (Section 5 )
Table 1. Comparison of different underwater ranging techniques. Ranges are those reported or tested in the cited work.
Component
Source
Typically owned?
Share of cost
Dive camera
Consumer retailer
Yes, by underwater photographers
45%
Underwater housing
Consumer retailer
Yes, with the camera
31%
Wide-angle wet lens
Consumer retailer
Sometimes
16%
Dive laser pointer
Dive retailer
Sometimes
7%
Laser mount
3D printed from files released with the paper
No, printed
1%
Table 2. Components of the FishSense Lite system. Every part is available through consumer channels, and none requires a quote or a specialist supplier. The mount is the only part that must be made rather than bought. Shares are of October 2026 retail prices for an equivalent unit, about US$1,300.
Figure 2. The principle of laser triangulation underlying FishSense Lite. (a) A pinhole camera projects a fish of length L at distance d to an inverted image of length p . (b) A laser mounted offset from the camera; points A and B are at different distances along the laser ray. (c) The same points in the image, sliding toward the beam’s vanishing point as distance grows.
Figure 3. A FishSense Lite frame. The white ring marks the laser dot; its position in the image gives the fish’s range. The amber marks are the snout and tail fork as the labeler placed them, and the length is the distance between those two points back-projected to the dot’s depth. An underwater photograph of a blue parrotfish with the laser dot and the two measured endpoints marked.
Figure 4. Laser ray model and ray intersection. (a) The laser, with origin ℓ and direction α , intersects the fish at P . (b) The camera ray and laser ray share the terminus P , forming an overdetermined system in λ1 and λ2 .
Figure 5. Field calibration target: a standard underwater dive slate with a chevron of black duct tape.
Cohort
Frames
Targets
Reference
Paired stereo dives
39
7 wild fish
stereo-imaging system
Pool
1,055
5 rigid models
tape or measuring board
Field
162
73 wild fish
none
Table 3. The three cohorts this paper evaluates. The angle experiment’s 1,428 frames are reported separately (Section 5.4 ).
Target
Length
Profile
Box
150 mm
flat
Purple Angelfish
192 mm
flat
Rainbow Trout
313 mm
thick
Grouper
360 mm
flat
Snook
455 mm
thick
Table 4. Rigid targets used in the pool. The flat targets match the planar assumption of Section 3.5 , while the thick targets carry the body depth of a real fish.
Figure 6. FishSense Lite against an underwater stereo-imaging system on the same seven individuals. Each orange diamond is one fish’s FishSense Lite length, the p90 of its frames; the blue points are those frames, 39 in all, between 3 and 11 per fish. The dashed line marks agreement, not truth; neither axis is a known length. Scatter plot of \fsllength against stereo-imaging length for seven fish, with a 1:1 line.
Species
Dive
Frames
Difference from stereo
Hogfish
1
11
+3.5%
Hogfish
2
3
+8.6%
Hogfish
1
3
−13.8%
Blue Parrotfish
1
10
+10.7%
Stoplight Parrotfish
1
6
+11.2%
Gray Snapper
2
3
−8.1%
Table 5. The seven fish measured by both instruments, from two dives under one morning calibration. Each difference is the FishSense Lite p90 minus the stereo length, as a percentage of the stereo length. Within each dive the differences span most of the full range (dive 1: −13.8 to +11.2% ; dive 2: −8.1 to +8.6% ), while the two dive means differ by 4.9 points.
Figure 7. Measured against known length over the pool cohort, 1,055 frames of five rigid targets. Diamonds are the per-target p90 . R2=0.999 about the 1:1 line on those five estimates, and 0.976 on the individual frames. Scatter plot of measured length against known length for five rigid targets.
Figure 8. Percent length error by target. Box and whiskers give the frame-level spread, the diamond the reported p90 . All 1,055 frames are shown, out to the worst at −30.6% . Box plot of percent length error for each of the five rigid targets.
Figure 9. How far a pool target’s p90 moves when estimated from n randomly drawn frames rather than all its frames, across the 15 sessions in which the Box or the Rainbow Trout was photographed at least 30 times. The shaded band holds 80% of draws. Below ten frames, the p90 is the longest frame. Line plot of the change in the p90 estimate against the number of frames per target, with an 80 percent band of random draws.
Figure 10. Each row is one pool session measured with its own calibration and, on the same frames, with the same unit’s next calibration. A neighboring calibration moves 21 of 22 sessions further from their targets’ known lengths and 16 of them past the 15% budget, by an amount that tracks how far the beam moved between the two calibrations. Dumbbell plot of each pool session's median length error with its own calibration and with the next calibration of the same unit, sorted by the latter.
Figure 11. Percent length error against the fish’s angle from perpendicular, from five sessions of one target stepped through 0∘ to 45∘ . Thin lines are per-session medians, black is the pooled median and interquartile range, dashed is cosθ−1 , and dotted is the 15% budget. Line plot of percent length error against fish angle for five sessions.
Figure 12. Median length error against laser range for the Rainbow Trout, a thick model 58.7 mm deep, and the Box, a flat face, in fixed range bins over the pool cohort. Shaded: ranges under 1 m. Line plot of median length error against laser range for the trout and the box, with ranges under 1 m shaded.
Figure 13. Length per animal by species over the seven dives, one point per animal at its p90 . The bar marks the median for species with three or more animals. Strip plot of per-animal length grouped by species.
Figure 14. Within-individual repeatability, wild fish against posed models. Each point is one wild individual with three or more frames, or one target photographed repeatedly in one pool session. Bars give the median and its 95% confidence interval. Strip plot of within-individual coefficient of variation for wild fish and pool targets.
Error source
Size (% of length)
Behavior
Section
Calibration scale between sessions
2.4 (SD)
random per session
5.4
Laser movement within a dive (two dives)
2.2 (SD)
random per dive
5.3
Labeling noise at 2 m
1.0 (SD)
random per frame
5.5
Few frames per fish
1.3 (SD)
random per fish
5.2
Pose within 15∘
up to 3.4
always shortens
5.4
Body depth
1.5 at 2 m, 5.0 at 0.58 m
always shortens
5.2 , 5.5
Table 6. The error sources measured in this paper, as magnitudes in percent of length. Each random term is a standard deviation: calibration scale, the spread of the scale errors of the five angle-experiment sessions ( −3.0 , +0.2 , −6.5 , −2.3 and −3.3% ; Section 5.4 ); laser movement within a dive, the largest change observed between a dive’s two ends, 3.7%, treated as the edge of a uniform distribution; labeling noise, the 0.9 px dot and 2.3 px endpoint errors of Section 5.5 for a 30 cm fish at 2 m; and few frames per fish, the spread a three-frame p90 adds over an all-frames p90 across the pool targets of Section 5.2 . These combine as independent errors, as the square root of the sum of their squares, to 3.6%. The last two rows always shorten a length, so they add rather than cancel.
Monitoring fish growth behavior provides relevant information about fish health in aquaculture and home aquariums. Yet, monitoring fish sizes poses different challenges, as fish are small and subject to strong refractive distortions in aquarium environments. Image-based measurement offers a practical, non-invasive alternative that allows frequent monitoring without disturbing the fish. In this paper, we propose a non-invasive refraction-aware stereo vision method to estimate fish length in aquariums. Our approach uses a YOLOv11-Pose network to detect fish and predict anatomical keypoints on the fish in each stereo image. A refraction-aware epipolar constraint accounting for the air-glass-water interfaces enables robust matching, and unreliable detections are removed using a learned quality score. A subsequent refraction-aware 3D triangulation recovers 3D keypoints, from which fish length is measured. We validate our approach on a new stereo dataset of endangered Sulawesi ricefish captured under aquarium-like conditions and demonstrate that filtering low-quality detections is essential for accurate length estimation. The proposed system offers a simple and practical solution for non-invasive growth monitoring and can be easily applied in home aquariums.
Clemens Seibold, Anna Hilsmann, Peter Eisert
Humboldt University of Berlin, Berlin, Germany · Fraunhofer HHI, Berlin, Germany
Accurate counts of fish populations are essential for monitoring ecosystems and making decisions for conservation, harvest limits, and river management. In many field deployments, these counts are obtained from sonar videos that experts must inspect manually, making large-scale monitoring slow, expensive, and difficult. Automated counting systems have been proposed as a scalable alternative, but so far this promise has been difficult to realize in deployment. Developing evaluations that better reflect real-world deployment conditions will help close this gap and drive improvements in counting performance in practice. To achieve this, we propose Caltech Fish Counting 2026 (CFC26), a new dataset with improved evaluation protocols that simulate deployment to seen and novel rivers, testing in-distribution and out-of-distribution performance. We expand the CFC22 dataset to eleven deployment locations across nine Pacific salmon river systems from Alaska to Northern California, capturing a broader geographical and ecological range of data. We additionally introduce two new evaluation metrics for counting, nMNE and nMANE, that mirror the operational workflow that experts use and help surface errors that frame-level detection metrics do not capture. Using these metrics, we reveal previously-concealed directional bias in upstream vs downstream counts, which would have direct and significant ecological implications if undetected. The dataset and code are publicly available on CaltechDATA and GitHub.
Madison Van Horn, Suzanne Stathatos, Sevan Brodjian +6
California Institute of Technology · Massachusetts Institute of Technology · California Institute of Technology Volunteer
Understanding the properties of fish populations is increasingly important. Analyzing fish length, width, and mass is integral to maintaining healthy stock in the aquaculture industry, while counting wild fish populations supports marine conservation and coastal reef restoration efforts. Deep learning-based segmentation models offer a promising means of obtaining such information at scale. Most of these models, however, require costly and time-consuming per-pixel segmentation labels for training, which can take up to two minutes per fish to annotate manually. To address this inefficiency, we propose a weakly supervised segmentation pipeline that requires only point-labeled data (one click per fish) to produce accurate segmentation masks. Our pipeline consists of four stages. First, a point-supervised tracker predicts point annotations for each frame. These points are then converted into motion-informed language expressions, which guide a language-conditioned segmentation model (LMPM++) to capture temporal consistency and produce candidate segmentation masks. Finally, optical flow and background subtraction are applied as post-processing filters to suppress noise from incorrect segmentations. Experiments on the DeepFish dataset show improvement in foreground mean intersection over union (mIoU) as compared to baseline models.