https://orcid.org/0000-0003-0469-740X
Figures
Abstract
In industrial societies, most fishery research institutes collect the phenotypic data of fish manually, which is time-consuming, labor-intensive, error-prone, and results in incomplete data. Considering their stress reaction and the natural body extension to collect the phenotypic data of fish quickly and accurately, channel catfish was used as the research subject and a deep-learning-based method was developed to explore their phenotypic data, i.e., body length, full length, head length, body height, tail handle width, tail handle height, and body thickness. First, this study applied two cameras and another device built into an image acquisition system to obtain images of fish in the water. We then adopted an Hourglass module network to position nine and ten key points on the top and side view images, building two key point fish skeletons. Finally, 3D coordinate transformation and scale parameters were employed to obtain the phenotypic data. Compared with the ground truth of the phenotypic fish data, our study achieved a 3.7% average relative error in terms of the full length, and an average 9.6% relative error for all seven types of phenotypic data applied. Furthermore, the average time required for the image processing measurements was approximately 1s.
Citation: Zhang X, Fang S, Li Y, Chen X, Huang F (2025) Point-based method for measuring the phenotypic data of channel catfish (Ictalurus punctatus). PLoS One 20(6): e0324158. https://doi.org/10.1371/journal.pone.0324158
Editor: Jay Richard Stauffer, Jr., Penn State University, UNITED STATES OF AMERICA
Received: October 24, 2024; Accepted: April 21, 2025; Published: June 5, 2025
Copyright: © 2025 Zhang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the article and its supporting information files.
Funding: This work was supported by the Zhejiang Public Welfare Technology Application Research Project of China (LGN21C190002) and the National Key Research and Development Program of China, "Innovation in Marine Agriculture and Freshwater Fisheries Science and Technology" Special Project, "Creation of New Germplasm and Breeding of New Varieties of Grass Carp with High Quality, Fast Growth, and Hemorrhagic Disease Resistance" (Grant No. 2023YFD2401600). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
In 1984, Channel Catfish (Ictalurus punctatus) was introduced to China from the US [1]. Due to its desirable traits for commercial cultivation, the cultivation area and total production of channel catfish in China increased significantly, reaching 248,000 metric tons and generating over 50 million US dollars in export income in 2014 [2]. Obtaining fish phenotypic data can help providing management advice and beneficial the economic growth of fishery enterprises. Furthermore, phenotypic data can be utilized for body-size-based sales strategies, enabling maximum economic returns by determining optimal harvesting times [3]. In addition to these advantages, phenotypic data serve as an indispensable reference in fishery breeding research, providing critical insights for scientific studies. Furthermore, phenotypic data also plays a significant role in fields such as fish taxonomy and systematics, as well as fish biology and ecology [4,5]. However, a traditional manual measurement method is time-consuming, labor-intensive, and subjective. Moreover, out-of-water measurements may hurt the measured fish. Laser measurements are expensive and impractical, and thus vision-based measurements have been considered. Computer vision is a non-destructive [6], fast [7], economic [8], consistent [9], objective, repeatable [10], quantitative [11], relatively effective, and powerful tool [12,13]. The application of computer vision to live-fish measurements is a challenging task. Owing to the high efficiency of computer vision measurement technology, this method is becoming increasingly popular [14]. Over the years, numerous researchers have been exploring how to obtain phenotypic data on fishes from digital images more accurately and efficiently. Traditional phenotypic data measurement methods are mostly out-of-water measurements, which can be divided into two categories: linear and non-linear. Recent studies are summarized as follows.
1.1 Linear measurement
When the fish’s body is approximately a straight line, the length of the body is equal to the length of the line segment in the image. The fish length can be calculated based on the relation between the line segment in the image and the image scale parameters [11]. The image scale parameters can be obtained through a square on the color plate [15]. The problem is how to obtain the line segment length on the image. Hsieh et al. [11] employed gradient-based edge detection combined with the Hough transform to identify the head and tail positions of tuna species, such as Bigeye Tuna (Thunnus obesus) and Albacore (Thunnus alalunga), by analyzing both sides of the color plate grid. In addition, Balaban et al. [16] used the bounding box to determine the width and length of the fish simultaneously, this method is only suitable for measuring the length of fish with straight bodies, and the categories of measured phenotypic data are relatively limited. To avoid interference from the external environment, Abdullah et al. [17] used computer vision technology and curvature information to extract the length parameters of the fish. Although this method can calculate full-length information, it relies on traditional image processing techniques and optical formulas, requiring a five-step calculation process that is cumbersome. Nelson et al. [18] developed a method for collecting fish length data using Electronic Monitoring (EM) systems. In this method, haddock (Melanogrammus aeglefinus) is placed horizontally on a measuring board during the measurement process, but only the length dimensions of the fish can be output.
1.2 Non-linear measurement
There is a large deviation in the fish length measurement based on linear measurement, many current studies on fish phenotypic data measurement employ nonlinear measurement methods. One of the methods is estimating the length by key points. Strachan [7] used the key point method to measure the length of the Haddock (Melanogrammus aeglefinus) on the conveyor. This study was based on the keypoint method to measure the length of the haddock on a conveyor belt. Under ideal constraint conditions, the measurement error was less than 1%. However, if the fish’s posture was random, the error increased to 3%. The measurement method proposed in this paper can only be measured out of water and has strict requirements for the fish’s posture. White et al. [19] used a binarized image to locate the direction of the fish’s body and cut the binarized body equally to obtain the midpoint through the perpendicular of this direction. The accuracy of this method is high, but if the angle of the fishtail is large, the final measurement error increases significantly. In other methods, Romero et al. [20] proposed to use of a third-order polynomial equation to fit the centerline of the rainbow trout’s body contour and then combine the Mahalanobis distance with a threshold to classify the fish. Miranda et al. [21] proposed to use the ratio of pixel changes between each frame of video images to determine the rainbow trout entering a specific shooting area, and then used a third-order polynomial equation to fit the centerline of the fish body contour to determine the length.
The current methods for measuring fish length are highly complex. However, they are prone to significant measurement errors when the fish curl their bodies. At the same time, the processes are cumbersome, and most methods are out-of-water measurements, which can cause a high-stress reaction.
1.3 Deep learning
With the development of the BP (backpropagation algorithm) [22], deep learning has developed rapidly, and CNN (convolutional neural network) [23] is being widely used in various image application fields. Tseng et al. [24] proposed to use a convolutional neural network to detect the positions of fish heads, fish tails, color plates, and then combined traditional image processing methods to calculate the actual value of the fish length. The detection of this method is more complicated, and the position of fish head and fishtail points is susceptible to environmental interference. Monkman et al. [25] proposed using R-CNN to detect the pre-examination position and the area of the color plate. They used the length of the largest detection bounding box as the predicted width of the fish’s body, combined with an ArUco marker [26] to predict the final length value. The positioning accuracy of the bounding box limits this method. Wang et al. [27] used image segmentation methods to measure body length and height of Large Yellow Croaker (Larimichthys crocea), with an error of less than 4%. The above three measurement methods based on deep learning are all performed out of water. To measure fish in water, an underwater stereo vision system was used to complete the measurement of fish body length [28,29].
From the above research, we can see that the number of phenotypic data measured by all research methods is small, and deep learning can avoid the artificial design of complex measurement algorithms. Based on this discovery, a key point positioning method based on deep learning is considered the preferred method in this study. The Hourglass Network is a model proposed by Newell et al. [30]. The cascaded hourglass network module achieves precise positioning of human key points. Compared to other methods, this deep network model based on an Hourglass Network module has a concise and easy-to-expand structure, showing an accurate and efficient performance. This study proposes a key point location algorithm based on the Hourglass Network, combining 3D transformation and scale parameters to obtain phenotypic data.
The main contributions of the paper are as follows:
- 1). Constructed a real-time collection device for fish top and side views, achieving unconstrained measurement of fish phenotypic data in water.
- 2). Designed the key point skeleton structure of catfish and created the corresponding dataset.
- 3). Established a deep learning detection model for fish body key points and implemented the determination of 7 phenotypic data of fish bodies based on three-dimensional coordinate transformation.
The material and methods are described in Section 2. Next, a deep learning model based on Hourglass Module, the image scale parameters, and the phenotypic data acquisition is displayed in Section 3. Results and discussion are described in Section 4. Finally, Section 5 conclude the advantages and disadvantages of the proposed method in this study.
2. Materials and methods
2.1 Phenotypic data on fish
The research subject applied in this study is the channel catfish. The measured phenotypic trait include the body length, full length, head length, body height, tail handle length, tail handle height, and body thickness. A specific example of each type of data is shown in Fig 1.
Ethics Statement
All procedures involving live channel catfish were conducted in accordance with the ethical guidelines and approved by the Laboratory Animal Welfare and Ethics Committee of the Zhejiang Academy of Agricultural Sciences (Approval No. 25ZALAS12). During the experiments, efforts were made to minimize stress and discomfort to the fish. After measurements, the fish were returned to their original aquaculture environment without harm.
2.2 Hardware device structure
The hardware device used in this study includes four strip light sources on the inside surface of the device, a slide track inside the device, and two four-megapixel cameras placed in a straight-up direction and toward the left direction. Fish boxes were used in this study, one side of which is transparent, and the other side is black. When measuring the phenotypic data, the fish box is filled with 3/4 volume of water before the fish is put in, and the fish box is then pushed into the device after being placed on the slide track. Based on Daheng’s industrial cameras (model: MER-503-20GC-P), we independently designed an image acquisition and testing device. The device structure is shown in Fig 2.
1 is the placement location of the side-view camera; 2 is the placement location of the top-view camera; 3 is the placement location of the fish storage box.
2.3 Data preparation
In this study, a deep learning network based on key point positioning is adopted. To better locate the coordinates of the key points of the fish’s body, a training dataset needs to be prepared before model training. Based on different images, we defined two key point skeletons for the side and top images. To measure seven sets of phenotypic data for the fish body, we defined labels for key points. Examples of keypoints for the fish side view and top view are as follows: a total of 10 keypoints are labeled in the side view, and 9 keypoints are labeled in the top view, arranged in order according to their positions. The label descriptions are shown in Table 1, and specific examples of keypoints in the figures are shown in Figs 3 and 4.
2.4 Model inference structure
The position of channel catfish in the water does not always maintain a straight line, most of the states are curved, and the phenotypic data on the fish cannot be accurately and effectively measured using a single camera. To this end, this study is based on computer vision and combines human pose estimation methods to train key point positioning network models based on the side and top images. During the model inference, the model outputs the coordinates of the key points of individual channel catfish in the side and top images. Based on these key point coordinates, the algorithm combines the 3D coordinate transformation and scale parameters to measure the final phenotypic data. The model inference structure is shown in Fig 5.
3 Theory/calculation
3.1 Mode and training parameter
The structure of our network is shown in Fig 6, which consists of two fourth-order hourglass modules [30]. An Hourglass module consists of residual modules. The main design of a residual module includes shortcut connections and identity mapping; the shortcut connections make the residual possible and deepen the identity mapping of the net, allowing a gradient explosion to be avoided. The topology of an Hourglass module is symmetric, and thus for each stage presented on the way down, there is a corresponding stage going up. In the down-sampling phase, features are extracted at each scale, whereasin the up-sampling phase, the original stage features and the features obtained from up-sampling from the previous stage are passed on to the next stage, the structure of which is shown in Fig 7.
The cascaded Hourglass network module achieves accurate human body key point positioning. Compared with other methods, this deep network model based on the Hourglass network module has a concise and easy-to-expand structure, showing an accurate and efficient performance. In this study, we use two fourth-order Hourglass networks to build a key point positioning system based on the top and side views. After determining the main network model, this study trains two models to obtain key point information and prepare for the subsequent phenotypic data acquisition.
The resolution of an input image is 2,448 × 2,048 and the ground truth heatmap resolution used for training in this study is 512 × 512. The Adam optimizer is applied when optimizing the mean square error loss function, with an initial learning rate of 1e-3. During the training and inference process, we use an NVIDIA RTX2080 SUPER graphics card with a 3.70 GHz Intel i7-8700K CPU.
3.2 Acquisition of scale parameters
The phenotypic data on fish can be calculated based on the relation between the line segment in the image and the image scale parameters. The scale parameters between the pixel length of the ruler and the actual length are obtained by the two cameras according to the proportional relationship, an image of which is shown in Fig 8.
3.3 3D coordinate transformation
Because fish in water do not always maintain a straight posture, it is necessary to use 3D coordinate transformation to accurately extract their length in 3D space.
In the 3D transformation shown in Fig 9, OA is the real length of the object in a 3D space, OA′ is the projection of the line segment OA on the plane YOZ, α is the angle between the line segment OA′ and the OY direction axis, and β is the angle between the projection of the line segment OA on the plane XOY and the direction axis OY. From this, the actual length of the OA can be calculated as follows:
3.4 Phenotypic data acquisition
To facilitate a better understanding of the subsequent calculation formulas, some midpoints composed of points on the side and top views, along with various functions, are added in this study. The descriptions of the points and functions are shown in Table 2.
The points obtained from the side view are comprehensively processed. The side view of the fish is divided into five segments, and the cosine of the angle between each segment vector and the horizontal axis is calculated. The final phenotypic data information is calculated based on the segmentation angle and proportional relationship obtained from the top view image.
The body thickness (BT) is calculated as:
The body height (BH) is calculated as:
The tail handle height (THH) is calculated as:
The head length (HL) calculation related to a 3D coordinate transformation is as follows:
The body length (BL) calculation is more complicated. First, the calculation involves calculating the pixel distance from the side-head point to the head-middle point.
Then, the calculation involves calculating the pixel distance from the head-middle point to small-middle point in the side image.
Finally, the calculation combines the proportional relationship and 3D coordinate transformation to obtain the body length.
The calculation of the full length (FL) involves the length of the tail and the calculated body length.
The shortest distance of the projection of vectors Vec (S9, S8) and Vec (S10, S8) on vector Vec (S12, S11) is used as the tail length.
The full length (FL) is calculated as follows:
The calculation of the tail handle width (THW) involves calculating the projection of vectors Vec (S12, S5) and Vec (S8, S12) on vector Vec (S12, S11) combined with the 3D coordinate transformation to obtain the final length.
3.5 Error evaluation
To better evaluate the accuracy of the algorithm proposed in this study for measuring of the phenotypic data of Ictalurus punctatus, current mainstream evaluation indicators, including the following three aspects are used.
The root mean square error (RMSE) represents the sample standard deviation among differences between estimated and measured values.
The mean absolute error (MAE) shows how close the prediction is in the final result:
Finally, the average relative error (MRE) indicates how large the length error is relative to the measured length:
4 Results and discussion
4.1 Positioning effect map
As shown in Fig 10, the measurement method proposed in this study can effectively locate the key points of the fish’s body, even when the body is bent. This provides a reliable basis for subsequent phenotypic data calculations.
4.2 Overall error calculation
In this study, we measured 20 channel catfish. The true values of the phenotypic data for each channel catfish were determined by an experienced researcher. Simultaneously, the predicted values obtained through computer-based measurements were recorded in Table 3. In the table, the left seven columns represent manual measurements, while the right seven columns represent measurements obtained using the method proposed in this study.
The manually measured data, obtained by experienced researchers, served as the standard reference. The computer-measured data were compared against these standard values, and the measurement errors for different phenotypic data were calculated using the three error calculation methods described earlier. The results are summarized in Table 4.
The mean relative error statistics are shown in Fig 10.
The mean absolute error statistics are shown in Fig 11.
In this experiment, the total length of channel catfish ranged from 12 to 17 cm. As shown in Table 4, the average relative error for total length using the proposed measurement method is approximately 3.7%, and the average absolute error is 0.5 cm, indicating high accuracy. However, a larger error was observed in the measurement of tail handle height, with a mean absolute error of 0.25 cm and a mean relative error of 26.4%, which is relatively high.
We believe that the measurement error of the tail handle height can be summarized as follows:
- 1). For channel catfish with a total length of 12–17 cm, the tail handle height is generally no more than 20 mm. A bias of 1 mm can result in a relative error of at least 5%.
- 2). The tail stem is relatively flat in some regions, making it difficult to determine precise positions, which introduces errors in manual measurements.
- 3). Errors were caused by a manual annotation.
- 4). The point coordinates, obtained from 128x128 feature maps, were upscaled by a factor of 16 to obtain the predicted coordinates in the original resolution. This caused an error of at least 0.17 cm in all phenotypic data.
Thus, errors in the scale of a few millimeters can lead to a significant average relative error.
5 Conclusions
In this paper, a method was proposed for measuring fish phenotypic data based on deep learning. This method applies phenotypic data measurements based on a key point detection, which uses seven phenotypic variables and achieves a high measurement accuracy. The measurement method proposed in this paper achieves an average absolute error accuracy of 0.5167 cm at full length. In addition, the body length, body height, body thickness, tail handle length, tail handle height, and head length were measured with relatively good accuracy. However, a large error may occur in measuring the tail handle length. We will explore more effective ways to reduce this error in future studies. It is worth noting that although the research subject applied in this study is channel catfish, the measurement method proposed herein can be applied to the measurement of phenotypic data of other types of fish.
Novelties of the study: 1) In-water measurement. Traditional methods for measuring fish phenotype data are typically conducted out of water. This study developed a hardware measurement device capable of measuring fish phenotype data in water, significantly reducing the stress response of fish and improving animal welfare; 2) Dataset construction. Based on the phenotype data of the target fish, this paper designed the skeletal structures for the side and top views of the fish body. Key points on the fish body were annotated, and a comprehensive image dataset was constructed to support deep learning model training; 3) Comprehensive phenotype measurement. Traditional methods are often limited to measuring body length and width. This study expands the scope by simultaneously measuring seven sets of phenotypic data, providing a more comprehensive characterization of fish morphology.
Shortcomings of the method: 1) Measurement errors. Experimental data indicate that the relative errors for tail stem width and thickness measurements are relatively large. As discussed in Section 4.2, the width and thickness of the tail are usually measured in a deformed state; 2) Hardware Limitations. The resolution of the industrial cameras used for capturing fish images may limit measurement accuracy. While increasing camera resolution could improve results, it would also raise equipment costs. A balance between cost and measurement accuracy must be carefully considered to meet application requirements. Based on the constructed fish phenotype measurement device, the experimental computer can determine 7 sets of fish phenotype data within 1 second, meeting the demand for a large amount of phenotype data in aquatic animal breeding and optimization, which is beneficial for fisheries experts to combine with genes to cultivate new varieties. Although the research object of this experiment is catfish, the device can be adapted in the future by adjusting the detection model for key points on fish bodies. This adjustment would enable the determination of phenotype data for other aquatic animals, thereby meeting the needs of breeding professionals in a wider range of fields.
Supporting information
S1_Fig. Top-view and side-view photographs of 20 channel catfish.
https://doi.org/10.1371/journal.pone.0324158.s001
(ZIP)
References
- 1.
Chen SJ. Studies on molecular genetics of Channel catfish (Ictalurus punctatus) germplasm in China [Master’s thesis]. Dalian: Liaoning Normal University; 2008. Available from: https://cdmd.cnki.com.cn/Article/CDMD-10165-2009122225.htm
- 2. Zhong L, Song C, Chen X, Deng W, Xiao Y, Wang M, et al. Channel catfish in China: historical aspects, current status, and problems. Aquaculture. 2016;465:367–73.
- 3.
Naiberg A, Little JJ. A unified recognition and stereo vision system for size assessment of fish. In: Proceedings of 1994 IEEE workshop on applications of computer vision; 1994 Dec 5–7; Florida, USA. IEEE; 1994.
- 4. Keat-Chuan Ng C, Aun-Chuan Ooi P, Wong WL, Khoo G. A review of fish taxonomy conventions and species identification techniques. J Surv Fish Sci. 2017;4(1):54–93.
- 5. Fu G, Yuna Y. Phenotyping and phenomics in aquaculture breeding. Aquaculture Fisheries. 2022;7(2):140–6.
- 6. Dunbrack RL. In situ measurement of fish body length using perspective-based remote stereo-video. Fish Res. 2006;82(1–3):327–31.
- 7. Strachan N. Length measurement of fish by computer vision. Comput Electron Agric. 1993;8(2):93–104.
- 8. St. John J, Russ GR, Gladstone W. Accuracy and bias of visual estimates of numbers, size structure and biomass of a coral reef fish. Mar Ecol Prog Ser. 1990;64:253–62.
- 9. Cocito S, Sgorbini S, Peirano A, Valle M. 3-D reconstruction of biological objects using underwater video technique and image processing. J Exp Mar Biol Ecol. 2003;297(1):57–70.
- 10. Balaban MO, Chombeau M, Cırban D, Gümüş B. Prediction of the weight of Alaskan pollock using image analysis. J Food Sci. 2010;75(8):E552-6. pmid:21535495
- 11. Hsieh CL, Chang HY, Chen FH, Liou JH, Chang SK, Lin TT. A simple and effective digital imaging approach for tuna fish length measurement compatible with fishing operations. Comput Electron Agric. 2011;75(1):44–51.
- 12. Harvey E, Fletcher D, Shortis M. A comparison of the precision and accuracy of estimates of reef-fish lengths determined visually by divers with estimates produced by a stereo-video system. Fish Bull. 2001;99(1):63–71.
- 13. Gümüş B, Balaban M, Ünlüsayın M. Machine vision applications to aquatic foods: a review. Turk J Fish Aquat Sci. 2011;11(1):171–81.
- 14. Karplus I, Alchanatis V, Zion B. Guidance of groups of guppies (Poecilia reticulata) to allow sorting by computer vision. Aquacult Eng. 2005;32(3–4):509–20.
- 15. Chang S-K, Lin T-T, Lin G-H, Chang H-Y, Hsieh C-L. How to collect verifiable length data on tuna from photographs: an approach for sample vessels. ICES J Mar Sci. 2009;66(5):907–15.
- 16. Balaban MO, Unal Sengör GF, Gil Soriano M, Guillén Ruiz E. Using image analysis to predict the weight of Alaskan salmon of different species. J Food Sci. 2010;75(3):E157–62. pmid:20492289
- 17.
Abdullah N, Shafry M, Rahim M, Amin IM. Measuring fish length from digital images (filed). In: Proceedings of the 2nd international conference on interaction sciences: information technology, culture and human; 2009. p. 38–43.
- 18. Nelson L, McElroy W, Jones A, Maynard G. How the collection of fish length data using electronic monitoring video review measures up to current methods used on a fishery-independent survey. J Northw Atl Fish Sci. 2024;55:1–10.
- 19. White DJ, Svellingen C, Strachan NJC. Automated measurement of species and length of fish by computer vision. Fish Res. 2006;80(2–3):203–10.
- 20.
Romero M, Miranda JM, Montes HA, Acosta JC. A statistical measuring system for rainbow trout. In: Proceedings of the international conference on image processing, computer vision, and pattern recognition (IPCV); 2014 Jul 21–24; Las Vegas, USA. Las Vegas: The Steering Committee of the World Congress in Computer Science; 2014. p. 1.
- 21. Miranda J, Romero M. A prototype to measure rainbow trout’s length using image processing. Aquacult Eng. 2017;76:41–9.
- 22. LeCun Y, Boser B, Denker JS, Henderson D, Howard RE, Hubbard W, et al. Backpropagation applied to handwritten zip code recognition. Neural Comput. 1989;1(4):541–51.
- 23. Lecun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proc IEEE. 1998;86(11):2278–324.
- 24. Tseng CH, Hsieh CL, Kuo YF. Automatic measurement of the body length of harvested fish using convolutional neural networks. Biosyst Eng. 2020;189:36–47.
- 25. Monkman GG, Hyder K, Kaiser MJ, Vidal FP. Using machine vision to estimate fish length from images using regional convolutional neural networks. Methods Ecol Evol. 2019;10(12):2045–56.
- 26. Garrido-Jurado S, Muñoz-Salinas R, Madrid-Cuevas FJ, Marín-Jiménez MJ. Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognit. 2014;47(6):2280–92.
- 27. Wang Y, Wang J, Xin R, Ke Q, Jiang P, Zhou T. Application of computer vision in morphological and body weight measurements of large yellow croaker (Larimichthys crocea). J Fish China. 2023;47(1):1–10.
- 28. Deng Y, Tan H, Tong M, Zhou D, Li Y, Zhu M. An automatic recognition method for fish species and length using an underwater stereo vision system. Fishes. 2022;7(6):326.
- 29. Zhou M, Shen P, Zhu H, Shen Y. In-water fish body-length measurement system based on stereo vision. Sensors (Basel). 2023;23(14):6325. pmid:37514620
- 30.
Newell A, Yang K, Deng J. Stacked hourglass networks for human pose estimation. In: European conference on computer vision; 2016 Oct 11–14. Amsterdam, Netherlands: Springer; 2016.