Methods and systems for non-invasive determination of animal mass

Abstract

Embodiments described herein relate to methods and systems for determining animal mass. The method can include receiving, by a processor of a computing device, video data representing observation of at least one animal, and executing, by the processor, a neural network architecture. Hie neural network architecture can be configured to receive an input video frame extracted from the video data, to generate an ellipse description of the at least one animal based upon the input video frame, the ellipse description being defined by predetermined ellipse parameters, and to estimate, based on the ellipse description, a mass value of the at least one animal.

Description

Attorney Docket No.23-1788-WO Methods and Systems for Non-Invasive Determination of Animal Mass CROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 615,843, filed December 29th, 2023, the contents of which are hereby incorporated by reference. STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with United States government support under Grant Nos. DA051235, DA048634, AG078530, and NS078795, awarded by the National Institute of Health (NIH). The United States government has certain rights in the invention. BACKGROUND

[0003] Unless otherwise indicated herein, the materials described in this section are not prior art to the claims in this application and are not admitted to be prior art by inclusion in this section.

[0004] Body mass is a primary measure of health and disease in humans. For example, body mass index is a key measure of metabolic health, and is one of the oldest and most widely used metrics in modern medicine. Changes in body mass indicate the function of many systems, including metabolic, cardiac, and the psychiatric, and are often the primary symptom of disease onset.

[0005] Rodents, particularly mice, are commonly used to model human diseases, carry out preclinical studies, and investigate disease pathologies. Over 95% of disease research that involves animal models is conducted with mice.

[0006] As in humans, changes in body mass in mice are predictive of health, particularly when sudden changes occur. The Institutional Animal Care and Use Committee(IACUC) includes significant body mass loss in its “human intervention” guide, a set of standardized criteria that call for veterinary intervention or euthanasia of subjects. Loss of body mass greater than 20% relative to baseline or to matched controls is a common justification for such intervention. Outside of ethical care, body mass is also an important feature collected during preclinical, metabolic, cardiovascular, and neurobiological studies. SUMMARY

[0007] In rodent studies, the current best method for determining the mass of the animal requires manually weighing the animal on a balance which has several consequences. For example, direct handling of the animal induces physiological stress responses, alters immune responses, and can have confounding effects on studies. Additionally, the animal’s mass measurement is static: body mass is typically measured every few days during the animal’s rest phase, depending on the protocol. Weight is often visually assessed during daily health checks by caretakers, and mass is manually measured only if the animal looks unhealthy or has lost weight. This can be subjective and can lead to variable treatment of the animal. Thus, manual methods are not amenable to continuous assessment, and rapid mass changes can be missed. These effects of human handling and static periodic measures of mass can weaken the validity and reproducibility of mouse experiments. Thus, a noninvasive and continuous method of monitoring animal mass would have great utility in multiple areas of biomedical research, including accurate measurement of animal mass over time to improve animal models of disease and to improve animal welfare.

[0008] Some noninvasive methods to determine animal mass have used computer- vision based approaches. Some such methods for determining animal mass in farming contexts use a multitude of visual metrics, including top-down silhouette area, eccentricity, perimeter, body length, and body width, and often use a wide range of modeling approaches, typically some form of linear regression or some type of neural network applied to predictmass directly from the image. Depth cameras have also been utilized for a better 3D representation of the animal, though the accuracy of this method can vary based on the amount of cameras used and thus can be challenging to scale.

[0009] Other noninvasive methods to determine animal mass have been implemented using highly engineered cages with balances and compartments. Although these methods alleviate animal handling issues and provide continuous mass measurement, the highly engineered cage designs require modification of housing conditions and are currently limited to singly housed animals. Thus, these methods can be challenging to scale and practically implement.

[0010] The embodiments described herein provide methods and systems for determining animal mass. The method applies computer vision and statistical modeling techniques to identify an animal within video data and determine animal mass based on this information. Such implementations of visual determination of animal mass using video enables noninvasive and continuous monitoring and can improve animal welfare and preclinical studies. Additionally, such implementations have the potential to expand weight assessment to novel environments and for continuous assessment. Additionally, these implementations can be easily incorporated into existing computer vision-based monitoring systems for routine use, and is suitable for multiple environments, both within and outside home cages, and can be scaled for continuous monitoring of mass in multiple animals (e.g., by performing example embodiments multiple times for each animal). Therefore, this approach is potentially highly generalizable and scalable. Mass determination as described in the embodiments herein is thus highly accurate and precise and can be further developed for various real-world applications.

[0011] This approach provides an advantage over existing industrial farming approaches as well. Errors in visual mass determination introduced by animal posture arelargely an issue of the animal not standing in an optimal location and / or posture and thus their silhouettes are constantly changing. For this reason, mass determination in farming contexts is often an easier task, as farm animals, such as cattle, tend to have rigid postures and relatively constant silhouettes compared to smaller animals such as mice. Therefore, these farming techniques cannot be as easily applied to smaller animals. Indeed, some farming applications deal with the prediction of carcass mass, where posture is even less of a factor.

[0012] Smaller animals, such as mice, on the other hand, are highly exible, have much smaller but highly deformable bodies, and quickly change shape. This deformability is a direct result of animal behavior and is subject to individual differences in genetics. Thus, the deformability of rodents is a major challenge to visual weight assessment. The embodiments herein combat this challenge by combining several visual metrics, including area and eccentricity, and applying multiple linear regression.

[0013] Accordingly, a first example embodiment involves a method for determining animal mass. The method can include receiving, by a processor of a computing device, video data representing observation of at least one animal, and executing, by the processor, a neural network architecture. The neural network architecture can be configured to receive an input video frame extracted from the video data, to generate an ellipse description of the at least one animal based upon the input video frame, the ellipse description being defined by predetermined ellipse parameters, and to estimate, based on the ellipse description, a mass value of the at least one animal.

[0014] A second example embodiment involves a system for determining animal mass. The system can include a data storage device that maintains video data representing observation of at least one animal. The system can also include a processor configured to receive the video data from the data storage device implement a neural network architecture configured to perform several operations. The operations can include receiving an input videoframe extracted from the video data, to generate an ellipse description of the at least one animal based upon the input video frame, the ellipse description being defined by predetermined ellipse parameters, and to estimate, based on the ellipse description, a mass value of the at least one animal.

[0015] A third example embodiment involves a non-transitory machine-readable medium storing instructions. The instructions, when executed by at least one processor of at least one computing system, causes the system to perform several operations. The operations can include receiving an input video frame extracted from the video data representing observation of at least one animal, to generate an ellipse description of the at least one animal based upon the input video frame, the ellipse description being defined by predetermined ellipse parameters, and to estimate, based on the ellipse description, a mass value of the at least one animal.

[0016] These, as well as other embodiments, aspects, advantages, and alternatives, will become apparent to those of ordinary skill in the art by reading the following detailed description, with reference where appropriate to the accompanying drawings. Further, this summary and other descriptions and figures provided herein are intended to illustrate embodiments by way of example only and, as such, that numerous variations are possible. For instance, structural elements and process steps can be rearranged, combined, distributed, eliminated, or otherwise changed, while remaining within the scope of the embodiments as claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Example embodiments should become apparent from the following description, which is given by way of example only, of at least one preferred but non-limiting embodiment, described in connection with the accompanying figures.

[0018] Figure 1 depicts an example computing system, in accordance with example embodiments.

[0019] Figure 2 depicts an overview of a process for determining animal mass from video data, in accordance with example embodiments.

[0020] Figure 3A depicts two images of a mouse and a graph of variation in area of the mouse, in accordance with example embodiments.

[0021] Figure 3B depicts two images of a mouse and a graph of variation in area of the mouse, in accordance with example embodiments.

[0022] Figure 3C depicts two images of a mouse and a graph of variation in area of the mouse, in accordance with example embodiments.

[0023] Figure 3D depicts two images of a mouse and a graph of variation in area of the mouse, in accordance with example embodiments.

[0024] Figure 4 depicts a graph of relative standard deviations across a variety of different mouse strains, in accordance with example embodiments.

[0025] Figure 5A depicts a graph of True Mass versus Predicted Mass, in accordance with example embodiments.

[0026] Figure 5B depicts a graph of True Mass versus Predicted Mass, in accordance with example embodiments.

[0027] Figure 5C depicts a graph of mean absolute error for several different models, in accordance with example embodiments.

[0028] Figure 5D depicts a graph of the coefficient of determination for several different models, in accordance with example embodiments.

[0029] Figure 6A depicts a graph of True Mass versus Predicted Mass for male and female mice, in accordance with example embodiments.

[0030] Figure 6B depicts a graph of True Mass versus Predicted Mass for several different mouse strains, in accordance with example embodiments.

[0031] Figure 7A depicts a table containing the mean observed and predicted masses for a variety of different mouse strains, in accordance with example embodiments.

[0032] Figure 7B depicts a bar graph of the relative standard deviations for a variety of different mouse strains, in accordance with example embodiments.

[0033] Figure 7C depicts a graph of relative standard deviations for observed mass versus the relative standard deviations for predicted mass for a variety of different mouse strains, in accordance with example embodiments.

[0034] Figure 8A depicts a comparison of the performance of several different models on prediction of mass on data from a longitudinal experiment, in accordance with example embodiments.

[0035] Figure 8B depicts a scatterplot comparison of predicted and observed weights from a longitudinal experiment, in accordance with example embodiments.

[0036] Figure 8C depicts data from several different test subjects from a longitudinal experiment, in accordance with example embodiments.

[0037] Figure 9A depicts a video analysis pipeline, in accordance with example embodiments.

[0038] Figure 9B depicts results of a multi-day tracking experiment, in accordance with example embodiments.

[0039] Figure 9C depicts results of a multi-day tracking experiment, in accordance with example embodiments.

[0040] Figure 10 depicts a flow chart of an example method.DETAILED DESCRIPTION

[0041] Example methods and systems are described herein. Any example embodiment or feature described herein is not necessarily to be construed as preferred or advantageous over other embodiments or features. The example embodiments described herein are not meant to be limiting. It will be readily understood that certain aspects of the disclosed systems and methods can be arranged and combined in a wide variety of different configurations, all of which are contemplated herein.

[0042] Furthermore, the particular arrangements shown in the figures should not be viewed as limiting. It should be understood that other embodiments might include more or less of each element shown in a given figure. In addition, some of the illustrated elements may be combined or omitted. Similarly, an example embodiment may include elements that are not illustrated in the figures.

[0043] Additionally, for clarity, exemplary embodiments of systems and corresponding methods for video capture of one or more animals and mass determination of one or more animals are discussed herein in the context of small rodents, such as mice. However, the disclosed embodiments can be employed and / or adapted for video capture of and / or determining the mass of other animals without limit or restriction. Example Computing Systems

[0044] Figure 1 depicts a simplified block diagram of an example computing device 100. The computing device 100 can in certain embodiments be configured to perform and / or can perform one or more acts and / or functions, such as those described in this disclosure. The computing device 100 can in certain embodiments include various components, such as a processor 102, a data storage unit 104, a communication interface 106, and / or a user interface 108. Each of these components can be connected to each other via a connection mechanism 110.

[0045] In this disclosure, the term “connection mechanism” means a mechanism that facilitates communication between two or more components, devices, systems, or other entities. A connection mechanism can be a relatively simple mechanism, such as a cable or system bus, or a relatively complex mechanism, such as a packet-based communication network (e.g., the Internet). In some instances, a connection mechanism can include a non- tangible medium (e.g., in the case where the connection is wireless).

[0046] The processor 102 can include a general-purpose processor (e.g., a microprocessor) and / or a special-purpose processor (e.g., a digital signal processor (DSP)). The processor 102 can execute program instructions included in the data storage unit 104 as discussed below.

[0047] The data storage unit 104 can include one or more volatile, non-volatile, removable, and / or non-removable storage components, such as magnetic, optical, and / or flash storage, and / or can be integrated in whole or in part with the processor 102. Further, the data storage unit 106 can take the form of a non-transitory computer-readable storage medium, having stored thereon program instructions (e.g., compiled or non-compiled program logic and / or machine code) that, upon execution by the processor 102, cause the computing device 100 to perform one or more acts and / or functions, such as those described in this disclosure. These program instructions can define, and / or be part of, a discrete software application. In some instances, the computing device 100 can execute program instructions in response to receiving an input, such as an input received via the communication interface 106 and / or the user interface 108. The data storage unit 104 can also store other types of data, such as those types described in this disclosure.

[0048] The communication interface 106 can allow the computing device 100 to connect with and / or communicate with another entity, such as another computing device, according to one or more protocols. In one example, the communication interface 108 can bea wired interface, such as an Ethernet interface. In another example, the communication interface 108 can be a wireless interface, such as a cellular or Wi-Fi interface. In this disclosure, a connection can be a direct connection or an indirect connection, the latter being a connection that passes through and / or traverses one or more entities, such as a router, switch, or other network device. Likewise, in this disclosure, a transmission can be a direct transmission or an indirect transmission. In some embodiments, the communication interface 106 can allow the computing device 100 to receive video data from a video or image capture device, for example a camera.

[0049] The user interface 108 can include hardware and / or software components that facilitate interaction between the computing device 100 and a user of the computing device 100, if applicable. As such, the user interface 108 can include input components such as a keyboard, a keypad, a mouse, a touch-sensitive panel, and / or a microphone, and / or output components such as a display device (which, for example, can be combined with a touch- sensitive panel), a sound speaker, and / or a haptic feedback system.

[0050] The computing device 100 can take various forms, such as a workstation terminal, a desktop computer, a laptop, a tablet, and / or a mobile smartphone. Additionally, as used herein, “mobile computing device” describes computing devices that are highly mobile (including a laptop, a tablet, and / or a mobile phone), as well as computing devices that are not as mobile (including a desktop computer, etc.). In a further aspect, the features described herein can involve some or all of these components arranged in different ways, including additional or fewer components and / or different types of components, among other possibilities.Method for Determining Animal Mass

[0051] Figure 2 depicts an overview of the method for determining animal mass from video data in the form of a process 200. In some embodiments, video data can be collected by a camera connected to the computing device 100 via the communication interface 106.

[0052] In summary, the process 200 involves collecting video data from a top-down perspective in an open eld. Each frame is segmented for the mouse, which describes the animal’s size. The segmented image is used to t an ellipse that describes the approximate posture of the animal and is then adjusted with covariates to improve the modeling and mass prediction.

[0053] Video data can be collected from a top-down perspective in an open environment that allows for clear observation of a mouse or other animal. In the embodiments described in the figures, the video data was recorded at 30 frames per second (fps), has an 8-bit monochrome depth, is of 55 minutes of length, and has a video resolution of 480 x 480 pixels. In other embodiments, different framerates, color depths, lengths, and resolutions can also be used.

[0054] The environment depicted in the embodiments described in the figures is a 52 x 52 x 23 cm arena comprised of white PVC plastic floors and grey PVC plastic walls. Each arena was illuminated by an LED light producing between 200 and 600 lux of light. The camera was mounted approximately 100 cm above each arena, with the zoom setting set to 8 pixels / cm. In other embodiments, other camera heights and zoom settings can also be used. Frame Segmentation

[0055] Each raw frame 202 of the video data is fed into a neural network 204. The neural network then segments each frame to determine if a mouse is present, and if so, the region of the frame where the mouse is located is set apart for further processing. The operation of a neural network (e.g., neural network 204) involves layers of interconnectedprocessing units, known as neurons. The neural network can be trained on datasets of different types of information from diverse sources, thereby enabling the neural network to learn to recognize a wide array of structures and patterns within inputs. The training process involves adjusting the weights of the connections between neurons using algorithms such as backpropagation, in conjunction with optimization techniques like stochastic gradient descent, to minimize the difference between the neural network’s output and expected output.

[0056] In some embodiments, this segmentation process can be performed by a segmentation network, which can be a deep neural network. A deep neural network is a type of neural network with one or more “hidden” layers of neurons between the input and output layers. The segmentation network is trained on a diversity of mouse images and provides accurate recognition of mice within images and video frames. For each video frame, the segmentation network generates a segmentation mask, which are the pixels where a mouse is determined to be present. Then, the network outputs a segmented image distinguishing the “mouse pixels” from the rest of the image. In Figure 2, this is depicted on the segmented image 206, with the white pixels representing where the network identified a mouse to be within the raw frame 202.

[0057] Additionally, the segmentation network produces another metric for further use besides the segmented image 206, specifically a count of the “mouse pixels” that represent the area of the mouse in the video. For each video file, the median of the segmentation pixel measurement across each frame is calculated, designated Apx. This provides a single area measurement per-mouse per-video that can be used for further processing and modeling.

[0058] The segmented image 206 can then have an ellipse fit onto the mouse-pixel “blob” in order to approximate the posture of the mouse in question, whether it is facing forwards, backwards, and so on. This is represented in Figure 2 by the ellipse-fit image 208.Covariate Adjustments

[0059] In some embodiments, the method can adjust the image in response to several covariates, which are variables that can affect the observation of the primary variable (in this case, animal mass) but are not directly intended to be studied or measured. Illustrated in the covariates block 210 in Figure 2 are a few examples of such covariates: environmental, posture, and per-animal (e.g., genetic information such as strain, variety, and / or breed). At least some of these covariates can also be referred to as predetermined ellipse parameters, as they can affect how the ellipse is fit onto the segmentation mask.

[0060] Environmental covariates are those that affect or are a result of the image capture process. For example, in order to ensure standard calculation of animal area for the mass-estimation model to function properly, differences in camera lens zoom must be accounted for, even if efforts are made to ensure they are as close as possible. To account for this, the measured segmentation (square-pixel units) must be converted to real-world metric units (square cm).

[0061] In the embodiments described in the figures, the arenas used to capture the video data of the animals measure 52 x 52 cm. Thus, to convert the segmentation pixel measurement, the computing system 100 can make use of a corner detection deep neural network to derive a scaling factor from pixels to centimeters.

[0062] Following this, the segmentation pixel measurement Apxcalculated previouslycan be adjusted using the following question: / / , where Acm is the unit-converted area of the mouse in square cm, Apxis the area in square pixels, Lpxis the length of the corner edge in pixels, and Lcm is the length of the corner edge in cm. In the embodiments described in the figures, the arenas measure 52x52cm, and thus Lcm would equal 52cm. In other embodiments where the arena size varies, Lcmwould reflect that size.

[0063] The advantage of adjusting the measurement in this way is that it does not necessarily require knowledge of the full size of the arena, but simply the real-world distance between the corners of the arena.

[0064] Thus far, this disclosure has identified a metric, Acm, which captures the unit- converted segmentation area of a single mouse, normalized for differences in camera zoom. However, as discussed previously, mice are highly deformable depending on their behavior, and over a short time frame a mouse can scrunch, stretch, bend, and / or rear such that its segmentation area quickly changes. This is one example of a posture covariate that must be accounted for. As the Acm metric only reflects the median area, this metric does not account for the wide variation in area for individual mice over short time periods, which is a direct result of changing behavior and posture in the mice. This can thus affect its precision, and must be accounted for if the model is to remain useful.

[0065] Figures 3A-3D depict four different mice of approximately the same mass (25.1 g) from different strains, examples of the smallest and largest segmentation areas they were each calculated to have over the 55-minute range of the video data, and a graph of how their area deviated from the median over the course of the 55 minutes.

[0066] Figure 3A depicts mouse C57BL / 6J, who was measured as having a median segmentation area measurement of 1434 ± 153 pixels.

[0067] Figure 3B depicts mouse C57BL / 6NJ, who was measured as having a median segmentation area measurement of 1355 ± 196 pixels.

[0068] Figure 3C depicts mouse A / J, who was measured as having a median segmentation area measurement of 1386 ± 142 pixels.

[0069] Figure 3D depicts mouse BALB / cJ, who was measured as having a median segmentation area measurement of 1421 ± 166 pixels.

[0070] Figures 3A-3D show that the area of the mouse commonly deviated by ±40% (relative to the mean area) over the course of the video data. The high variation in area was due to deformability of the animal due to altered behavior. This demonstrates that different strains of mice have different activity characteristics, something that depends at least partly on genetics.

[0071] Since multiple images were captured for each animal, a mean or median segmentation area may be more informative. However, variance in this measurement can impact the precision of a prediction. This was investigated by exploring changes in the variation of segmentation area between the mice in Figures 3A-3D. For instance, a 25.4g C57BL / 6NJ mouse (as in Figure 3B) has an average area of 1355 ± 196 px, whereas a smaller 24.5g, A / J mouse (as in Figure 3C) has a larger area of 1386 ± 142 px. However, the variation in area is larger in C57BL / 6NJ than in A / J. A / J are a high anxiety strain with low ambulation, while C57BL6 / NJ have more bouts of activity. The larger segmentation area and lower variation in A / J is because it spends more time in the corner in a constant posture. This demonstrates that simple segmentation is inadequate to handle genetic diversity seen in laboratory mice, and a more sophisticated approach is needed.

[0072] Due to the variation in area measurement as a result of posture differences as described above, the neural network can also adjust for these posture covariates and accordingly adjust Acm, as represented in Figure 2 by the covariates block 210. This can be accomplished through normalizing the area based on geometric shape descriptors of the segmentation mask, such as eccentricity, aspect ratio, and elongation. In some embodiments, the speed of the mouse can also be used to normalize the area.

[0073] One method of adjusting Acm can take into account hyperbolic eccentricity ofthe fitted ellipse. Hyperbolic eccentricity is defined as / / , where w and l arewidth and length respectively.

[0074] Eccentricity, as well as the other geometric shape descriptors listed above, were each tested to determine which provided the best indicators of shape, the results of which are illustrated in Table 1 below. Table 1 – Posture Adjustment Metrics

[0075] Table 1 analyzes different posture adjustment metrics, using four example models: T1, T2, T3, and T4, each correcting the ellipse for eccentricity, aspect ratio, elongation, and speed. Each model is a single-variable linear regression, taking the product of Acmand the median of one posture metric (eccentricity, aspect ratio, elongation, or speed) as its input variable and predicting body mass. T1’s input is equivalent to Ae as described above, while the inputs of T2, T3, and T4 are analogous to Ae but for aspect ratio, elongation, and speed according to the respective metric formulas in the table.

[0076] In the Metric Formulas in the table, w and l are the width and length of the fitted ellipse, vx and vy are the x and y components of velocity (also representable as ), and each μijis a central image moment. A central image moment is a weighted average of image pixel’s intensities and is used in computer vision applications.

[0077] Several different error measurements were used to evaluate the factors, including mean absolute error (MAE), mean absolute percent error (MAPE), root-mean- squared error (RMSE), as well as the coefficient of determination (R2).

[0078] In each of the following definitions of the error measurements, y and are the true (observed) and predicted values from the model, respectively, while n is the number of values.

[0082] R2 ,

[0083] With hyperbolic eccentricity, T1 achieves an R2of 0.827 and MAPE of 7.328%, significantly better performance than models T2-T4. This led us to conclude that eccentricity is the most useful indicator of shape of the metrics tested. It was initially hypothesized that speed would be a good descriptor to correct posture by, since mice tend to take on a constant elliptic shape when they’re moving. However, speed performs poorly, resulting in a MAPE of 19.4% and R2value of only 0.002. It is hypothesized that this lack of quality is due to the proportion of time mice actually move quickly. Although mice do take on relatively constant postures while walking forward, this kind of motion takes up a relatively tiny fraction of total frames, and is likely strain specific (or variable based on other genetic information). Thus, speed can in certain embodiments work for certain strains, but can also not generalize well in genetically diverse populations.

[0084] Thus, model T1 had the lowest error across MAE, RMSE, and MAPE, while having the highest R2, demonstrating that correcting for the eccentricity covariate improves the performance of the method overall.

[0085] From this result, the "eccentric area" metric can be formalized as , where Acm is the median of unit-converted area and e is the median hyperboliceccentricity, both over all frames of the given video data. In full, / / / / , where Acm is the unit-converted area of the mouse in square cm, Apx is thearea in square pixels, Lpx is the length of the corner edge in pixels, Lcm is the length of the corner edge in cm, and where w and l are the width and length of the fitted ellipse respectively. Aethus functions as the video-level summary metric for the mass determination method.

[0086] To quantify the effect of correcting posture with eccentricity, Figure 4 compares the variation of Acm and Ae by comparing their respective Relative Standard Deviations (RSD) across a variety of different mouse strains. Specifically, Figure 4 depicts the relative standard deviations, or RSD, of the segmentation area Acmin the upper set on the chart compared to the RSDs for the adjusted area Ae in the lower set on the chart for 62 different mouse strains. As shown, the RSDs for Ae were consistently lower for each strain, demonstrating further that correcting segmentation area for eccentricity improves the performance of the method.

[0087] Strain is another important factor in variability, as some strains like the obesity model NZO / HILtJ have a low RSD of around 6%, whereas a wild-derived strain like WSB / EiJ has a RSD around 15%. The largest area RSD observed was about 18%, for the SWR / J strain. It is possible that these strain-level differences in RSD are largely due to differing activity levels and behavioral patterns between strains. After posture correction relative standard deviation across all strains is significantly reduced (Figure 4, in the lower boxes). The RSD, however, is still strain dependent, indicating there are still strain level effects on the precision of the prediction even after posture correction.

[0088] The above results demonstrate that per-animal covariates (such as the aforementioned strain) can also affect the mass of an animal, and thus the neural network can also take such covariates into account before determining a final visual mass prediction 212. These covariates can be known or not, depending on the context the measurement is being taken in, but in some embodiments these covariates can include the sex of the animal, the strain or breed, the age of the animal, and the arena or specific open field / environment that the animal was observed in. Results

[0089] This disclosure has provided several ways of the describing size calculated or adjusted from the segmented video image of an animal: the raw segmentation area in pixels Apx, the unit-converted area Acm, and the eccentric area Ae. Thus, to evaluate each of these measurements in combination with the additionally covariates described above, this disclosure presents six multiple linear regression models as depicted in Table 2 below.Table 2 – Model Definitions and Performance

[0090] Each model uses one of raw area (Apx), unit-converted area (Acm), or eccentric area (Ae) as its visual input variable, and different subsets of sex, strain, age, and arena (the specific open field / environment the mouse was measured in) as covariates. After building each model, a 50-fold cross validation was performed on a 70 / 30 training / testing split. Averaging the reported accuracy and error values over these 50 iterations ensures that the analysis is not biased by lucky or unlucky sampling.

[0091] The first model in Table 2, Base (M1), is a single variable linear regression that uses raw segmentation area Apx to predict body mass, which was used as a comparative baseline for the other models in Table 2. This model has a R2value of 0.767 and MAPE of 8.584% (Figure 4).

[0092] The second model, M2, is also a single variable regression, but between unit- converted area Acm and body mass. M2 performs better than the Base (M1) model, increasing accuracy and decreasing error.

[0093] In M3 a second variable, arena, was added, referring to the particular arena a mouse was tested in. M3 performs slightly better than M2 by all measures.

[0094] M4 introduces eccentricity, which is combined with unit converted area to be one variable Ae, as previously described; M4 has two variables: Aeand arena. M4 performs substantially better, with R2of 0.822 and MAPE of 7.566%.

[0095] In a practical case of assuming minimal information, one can use M4 instead of M1, M2, or M3, because Apx, Acm, and Ae are visual metrics, and the particular arena is inherent to the experiment. One can in measure these variables and would have no reason not to, thus not using M4 in this case would decrease prediction performance for no experimental benefit. However, this disclosure presents each model herein because selectively adding and / or modifying variables demonstrates the performance benefit of each of unit conversion, arena identity, and eccentricity without confounding the variables. Additionally, likelihood- ratio tests confirm each model performs significantly better than the last, so one could use any model with confidence in some embodiments as the situation requires.

[0096] In some embodiments, the sex and age of the animal are known. Therefore, the fifth model in Table 2, Sex / Non-Genetic (M5), adds sex and age as separate variables, reflecting the differences in body composition between sexes and as a mouse ages. This again provides a boost in accuracy and reduction in error values.

[0097] In some embodiments, the strain of the mouse can also be known. Therefore, the sixth model, Full (M6), introduces the strain of the mouse as a variable. Here, the model takes into account an interaction term between strain and sex, because it was observed that while males are generally heavier than females, there are some strains for which the oppositeis true. Other genetic information, such as variety or breed, may also be used as a variable in some embodiments.

[0098] The Full (M6) model performs much better than the previous models M1-M5, achieving a mean R2of 0.92 and MAPE of 4.84% as shown in Table 2. Visualizing MAE and R2values of each model illustrates their relative improvements under cross-validation, which is depicted in Figures 5A-5D.

[0099] Figures 5A and 5B compare True Mass on the x-axis, as measured with a traditional scale, and Predicted Mass on the y-axis, as determined by the methods described herein. Also shown is the R2, MAE, RMSE, and MAPE for the relevant model. The 45º line indicates a perfect prediction (i.e. the True Mass equals the Predicted Mass).

[0100] Figure 5A depicts the above information for the Base Model (M1), as described above in Table 2, which only accounts for raw segmentation area in pixels.

[0101] Figure 5B depicts the above information for the Full Model (M6), as described above in Table 2, which accounts for the eccentric area Ae, arena, sex, age, and strain.

[0102] Figures 5C and 5D depict error and accuracy performance for all six of the models in Table 2 under 50-fold cross-validation.

[0103] Figure 5C depicts the MAE across the six models.

[0104] Figure 5D depicts the R2value across the six models.

[0105] As above noted, these figures demonstrate that each iteration of the models can increase performance, with the Full (M6) model standing out as particularly accurate.

[0106] The performance of the Full (M6) model is further set forth herein with relation to Figures 6A and 6B.

[0107] Generally, large differences are observed in mean body mass between sexes due to underlying changes in body composition. For most but not all mouse strains, males have higher percent fat than females. The performance of the full model (M6) was tested onmales and females. While females weigh less than males on average, the full model performs uniformly well across females and males, as the graph in Figure 6A shows. Differences in performance are slight: females and males have MAPE of 4.82% and 5.17%, respectively.

[0108] Performance of the full model (M6) was then tested across genetically diverse mice and measured accuracy and precision of the model across 44 classical inbred, 7 wild- derived inbred, and 11 F1 hybrid strains, with the results depicted in Figure 6B. Most strains are between 20-30 g with wild-derived strains like MSM / MsJ at the low end and obesity models such as NZO / HILtJ on the high end of the distribution. The model accurately predicted mean and variance across highly genetically diverse mice. The variance of predicted mass reflects the variance of true mass, and suggests that the model captures much but not all variation within strains.

[0109] The mean observed and predicted mass from the full model (M6) and standard deviation for each strain are depicted in the table in Figure 7A. Also depicted are the number of male and female mice included in the dataset from each strain reported.

[0110] Additionally, the full model (M6) can be retrained on the entire dataset, and present the observed and predicted relative standard deviation in mass for each strain in Figures 7B and 7C, demonstrating that real strain-level variation is captured in the model.

[0111] Specifically, Figure 7B depicts the relative standard deviations (RSD) of observed mass (upward-hatched boxes on the chart) and the mass predicted by the full model M6 (downward-hatched on the chart and trained on the full dataset) for each of the 62 mouse strains considered in the dataset.

[0112] Figure 7C, accordingly, depicts the RSD of mean observed mass on the x-axis compared to the RSD of mean predicted mass on the y-axis for each of the 62 mouse strains considered in the dataset. The black diagonal line is 45º, indicating exact correspondence between observed and predicted RSD.

[0113] This model can be used for inferring on any new data. Thus, it can be concluded that this model performs well across sex and diverse genetics. Applications of Visual Mass Estimation

[0114] The above methods and results generally concern the visual prediction of mass using genetically diverse mouse strains and ages that are tested once in the open field. However, other applications using the same and / or similar approaches are possible and contemplated herein, some of which are described as examples below. Longitudinal Tracking of Animal Mass

[0115] The above approach was tested to determine if it could consistently track the weight of the same mouse over time. The reasoning is that if a mouse suddenly lost weight or was under treatment that results in adverse effects, the non-invasive approach described herein would have diagnostic utility. The body mass of 16 mice were manually assessed and then tested in the 1-h open field over a 22-day period. Specifically, 4 C57BL / 6J females and 12 C57BL / 6J males were tested every day for 13 days, followed by an additional test day 9 days later (day 22).

[0116] Then, the previously trained M1–M6 models (described above) were applied to this data, and the performance was evaluated using quantitative and qualitative methods. These results are illustrated in Figures 8A-8C.

[0117] Figure 8A depicts a comparison of the performance (mean ± standard error) of the six models discussed above on prediction of mass on data from the longitudinal experiment. Figure 8B depicts a scatterplot comparison of predicted and observed weights of the 16 mice (R2= 0.87). Points on the dashed diagonal line indicate perfect predictions As shown in these figures, the full model, M6, performed best in terms of MAE (1.54 ± 0.08), RMSE (3.51 ± 0.33), MAPE (5:4% ± 0:25%), and R2= 0.87. Figure 8C depicts individual mouse data on each day. In particular, Figure 8C shows plots of predicted weights (dashed) toobserved weights (black) across each mouse for each day of the experiment. An (M) or (F) in the title of each panel indicates the sex of the animal. Qualitatively, the visual prediction of the mass of each of the 16 animals closely mirrored the manually measured mass, as shown in Figure 8C. These results demonstrate that the models herein perform well when tracking the mass of individual animals over time. Multi-Day Tracking with Multiple Animals in Home Environments

[0118] To further extend the applicability of these approaches, the performance of the models was tested in a long-term housing environment with multiple mice. This involved firstcollecting video data in a modi ed open- eld environment with bedding, food, and a waterLixit (water dispenser). This environment housed three female C57BL / 6J mice for 4 continuous days. The mass of each animal was measured manually at the beginning of the 4- day experiment and used as ground truth.

[0119] Then, 26 video clips of ~30 seconds or less were randomly selected for analysis from the active (dark) phase over a 4-day period. Since the visual mass models were trained on single mice, 8 clips were selected in which the mice were physically separated and distinct for analysis. This pipeline is illustrated in Figure 9A.

[0120] Each of these 8 clips is 10 to 20 seconds long and has a resolution of 800 x 800 px. The segmentation model used for single mice in the open eld was trained on diverse mice in a plain open eld without bedding, food, or water lixit. However, this open-field segmentation model generally did not perform well in the home environment with increased complexity, and thus a new segmentation model was needed. To predict multi-mouse segmentation, one approach applies the track anything model (TAM). TAM is a video extension of the segment anything model (SAM) that propagates segmentation predictions forward in time. The mice were initialized in a video using a manual keypoint prompt, which is typically 1 click per animal to indicate it as an object to track.

[0121] After this, the TAM can predict the segmentations of that mouse for the remainder of the video. The performance of this model was visually inspected to ensure thequality of the segmentation. This manual inspection of each frame of each video con rmedthat the identities of all mice remained consistent within a video. The identities of individual mice between videos were also manually matched. The resulting segmentation mask for each mouse was used for modeling using M6.

[0122] The results, illustrated in Figures 9B and 9C, indicated that M6 has very good performance in this new environment and, on average, closely follows the true mass over time in each video. In Figure 9B, the solid lines represent the median mass prediction in the time interval, while the true masses are shown as two black dashed lines (two mice had equal measured masses). The manual mass was measured at the start of the experiment. M6 also achieves a median error below 10% in all three mice: mouse 1 achieves a 3.15% mean error, mouse 2 achieves a 6.41% mean error, and mouse 3 achieves a 9.73% mean error, as shown in Figure 9C. Figure 9C depicts the MAPE of mass prediction for each of the three mice over the 8 videos. Each box shows the interquartile range (IQR) and the black whiskers reach the farthest point within 1.5 IQR of the end of the box.

[0123] Upon inspection of the videos, it was observed that the slightly higher error in mouse 3 appears to be due to its behavior. This mouse frequently jumped onto the walls and onto the lixit. Regardless, these error levels are comparable to the observed performance in the 1-h open eld, and thus it can be concluded that the approaches herein can be applied toward long-term home monitoring conditions for animals. Example Technical Improvements and Applications

[0124] The approach described in the foregoing can have several beneficial features. First, the approach uses an explainable machine learning approach. The first step of the approach is to generate segmentation masks for the mouse. Segmentation is highly flexibleand is one of the oldest and most established computer vision tasks particularly for biomedical images. Modern methods of segmentation improve upon traditional methods by using neural networks.

[0125] Segmentation neural networks can be trained with using much lower training data, and there exist pre-trained models that are intended for general purpose segmentation. Additionally, the creation of training data in different visual environments is made easier with new, more general segmentation foundation models such as the SAM and TAM (as discussed above) and self-supervised models such as DINOv2. These models appear to work well for segmentation of animals with minimal supervision as demonstrated herein.

[0126] The segmentation networks can also be modified. Some embodiments can include a segmentation step to enable observable detection for accuracy shifts. For example, if a mouse escapes the arena, the system can detect that no segmentation occurs and accordingly make no body mass prediction. After segmentation, the method applies techniques including unit transformations, normalization, and linear regression modeling. Thus, the method is flexible for adoption to new environments and organisms without the burden of creating large training data while producing explainable results.

[0127] Another possible advantage of the approach is the use of highly diverse mouse strains, with varied coat colors, sizes, and behaviors, for training the models. This ensures that the model can handle diverse visual and size distributions that can be seen in real world settings. Each of the six models described above have different applications.

[0128] Though the Full model has the highest accuracy and lowest error, it is not necessarily suitable for all situations. Practically, one can apply the Geometric, Non-Genetic, or Full model to an experiment, not M1, M2, or M3. While in some sense M1 is the most general of all the models, there’s no information in the Geometric model (M4) that is noteasily attainable. Its two variables, Ae and arena, are derived from the same base information in the video, not from any further information about the mouse itself.

[0129] For the same reasons, it would equally make sense to use M4 instead of M2 or M3. As mentioned above, it is still useful to present M1-M3, because they show the benefit of additions and modifications to the models.

[0130] With fine-tuning to the particular environment, the Geometric model (M4) can be applied to all open field videos of similar size and resolution with an individual mouse. However, it would be a rare mouse model experiment that intentionally did not take into account the sex and age of the mice, as these are often important variables.

[0131] The Sex and Non-Genetic model (M5), as its name suggests, would be very useful for experiments where collecting strain information is unfeasible or undesirable. A key example of this are experiments which use genetically diverse populations of mice, such as Diversity Outbred (DO) or Collaborative Cross (CC) populations. These types of populations are widely used to model genetic diversity, opposed to the standard inbred mouse strains, for example with application to high-resolution quantitative trait locus mapping (QTL), drug trials, drug exposure experiments, and disease experiments.

[0132] In these cases, it is not necessarily easy to determine the genetics of a given mouse, and genetic information can give little insight in the context of body mass prediction anyway, since the mouse can be unique or one of only a handful in the particular screen. This means the strain-free model (M5) is highly applicable to diverse mouse model experiments. The Full model, which includes genetic information, is readily applicable to any single- mouse open field experiment where strain is relevant and known. Examples include mouse colony management, where strain is known and highly relevant to body composition. The Full model would provide the highest accuracy monitoring of body mass in these areas. Furthermore, any classical mouse experiment with inbred strains that lends itself to havingone mouse in an open-field environment would benefit from this model, especially behavioral experiments where inducing anxiety potentially introduces bias.

[0133] For health monitoring, this style of visual prediction is only useful if its error falls well within the standards of humane intervention points, as to avoid false negative or false positive determinations of medically concerning body mass fluctuation. Following Institutional Animal Care and Use Committee (IACUC) Humane Intervention guidelines, it is necessary to intervene, typically with euthanasia, if a mouse loses more than 20% of its body mass compared to similar control animals.

[0134] As an example, if an inbred mouse lost 20% of its body mass during an experiment, the results in the foregoing disclosure indicate that the Full model (M6) could be used to predict that it had lost approximately 15.2-24.8% of its body mass, which is a small enough range to indicate that something is likely wrong with the mouse, and should be investigated further. In this way, this visual approach can serve well as a wide-screen diagnostic tool for health monitoring in its current state.

[0135] For applications that do not require adverse event detection, the sensitivity of the method described herein has added value. For instance, a drug or genetic manipulationthat leads to slight ( 5%) change in mass can be detected accurately. Although a 5% changein mass can seem low, in human clinical trials for weight loss drugs, the number of participants who achieve at least a 5% decrease in mass is the primary result that is reported. This level of sensitivity makes this a useful tool beyond health checks for preclinical animal studies. Example Method

[0136] Figure 10 is a flow chart illustrating an example method 1000.

[0137] At block 1002, the method 1000 can involve receiving, by a processor of a computing device, video data representing observation of at least one animal.

[0138] At block 1004, the method 1000 can involve receiving an input video frame extracted from the video data.

[0139] In some embodiments, the receiving of an input video frame extracted from the video data can be performed by a neural network architecture.

[0140] At block 1006, the method 1000 can involve generating an ellipse description of the at least one animal based upon the input video frame. In some embodiments, the ellipse description can be defined by predetermined ellipse parameters.

[0141] In some embodiments, the generating of the ellipse description can be performed by a neural network architecture.

[0142] In some embodiments, the neural network architecture can be a segmentation network that is configured to predict, on a pixel-wise basis, whether an animal is present in the input video frame. The segmentation network can then, in some embodiments, in response to predicting when an animal is present in the input video frame, output a segmentation mask.

[0143] In some embodiments, the segmentation mask can be based upon the pixel- wise prediction. The segmentation mask can also in some embodiments represent the animal in the input video frame.

[0144] The segmentation network can then, in some embodiments, fit the segmentation mask to an ellipse. In some embodiments, the ellipse comprises values characterizing the predetermined ellipse parameters.

[0145] In some embodiments, the predetermined ellipse parameters can include at least one of eccentricity, aspect ratio, and elongation. In some further embodiments, the ellipse description can include eccentric area, which in some embodiments can be based upon an environment-corrected area and the eccentricity. In some further embodiments, theenvironment-corrected area can be calculated based upon the segmentation mask and a measurement of an environment identified in the video data.

[0146] At block 1008, the method 1000 can involve estimating, based on the ellipse description of the at least one animal, a mass value of the at least one animal.

[0147] In some embodiments, the estimating a mass value of the at least one animal can be performed by a neural network architecture.

[0148] In some embodiments, estimating the mass value of the animal can include using at least one linear regression model and at least one multiple linear regression model. In some further embodiments, both models can be trained upon video data and animal-specific covariates. In some embodiments, the animal-specific covariates may include at least one of age, sex, and genetic information.

[0149] In some embodiments, the at least one animal can be a rodent.

[0150] The present disclosure is not to be limited in terms of the particular embodiments described in this application, which are intended as illustrations of various aspects. Many modifications and variations can be made without departing from its scope, as will be apparent to those skilled in the art. Functionally equivalent methods and apparatuses within the scope of the disclosure, in addition to those described herein, will be apparent to those skilled in the art from the foregoing descriptions. Such modifications and variations are intended to fall within the scope of the appended claims.

[0151] The above detailed description describes various features and operations of the disclosed systems, devices, and methods with reference to the accompanying figures. The example embodiments described herein and in the figures are not meant to be limiting. Other embodiments can be utilized, and other changes can be made, without departing from the scope of the subject matter presented herein. It will be readily understood that the aspects of the present disclosure, as generally described herein, and illustrated in the figures, can bearranged, substituted, combined, separated, and designed in a wide variety of different configurations.

[0152] With respect to any or all of the message flow diagrams, scenarios, and flow charts in the figures and as discussed herein, each step, block, and / or communication can represent a processing of information and / or a transmission of information in accordance with example embodiments. Alternative embodiments are included within the scope of these example embodiments. In these alternative embodiments, for example, operations described as steps, blocks, transmissions, communications, requests, responses, and / or messages can be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved. Further, more or fewer blocks and / or operations can be used with any of the message flow diagrams, scenarios, and flow charts discussed herein, and these message flow diagrams, scenarios, and flow charts can be combined with one another, in part or in whole.

[0153] A step or block that represents a processing of information can correspond to circuitry that can be configured to perform the specific logical functions of a herein-described method or technique. Alternatively or additionally, a step or block that represents a processing of information can correspond to a module, a segment, or a portion of program code (including related data). The program code can include one or more instructions executable by a processor for implementing specific logical operations or actions in the method or technique. The program code and / or related data can be stored on any type of computer readable medium such as a storage device including RAM, a disk drive, a solid state drive, or another storage medium.

[0154] The computer readable medium can also include non-transitory computer readable media such as computer readable media that store data for short periods of time like register memory and processor cache. The computer readable media can further include non-transitory computer readable media that store program code and / or data for longer periods of time. Thus, the computer readable media can include secondary or persistent long term storage, like ROM, optical or magnetic disks, solid state drives, or compact-disc read only memory (CD-ROM), for example. The computer readable media can also be any other volatile or non-volatile storage systems. A computer readable medium can be considered a computer readable storage medium, for example, or a tangible storage device.

[0155] Moreover, a step or block that represents one or more information transmissions can correspond to information transmissions between software and / or hardware modules in the same physical device. However, other information transmissions can be between software modules and / or hardware modules in different physical devices.

[0156] The particular arrangements shown in the figures should not be viewed as limiting. It should be understood that other embodiments can include more or less of each element shown in a given figure. Further, some of the illustrated elements can be combined or omitted. Yet further, an example embodiment can include elements that are not illustrated in the figures.

[0157] While various aspects and embodiments have been disclosed herein, other aspects and embodiments will be apparent to those skilled in the art. The various aspects and embodiments disclosed herein are for purpose of illustration and are not intended to be limiting, with the true scope being indicated by the following claims.

Claims

CLAIMS What is claimed is:

1. A method for determining animal mass, comprising: receiving, by a processor of a computing device, video data representing observation of at least one animal; executing, by the processor, a neural network architecture configured to perform operations comprising: receiving an input video frame extracted from the video data; generating an ellipse description of the at least one animal based upon the input video frame, wherein the ellipse description is defined by predetermined ellipse parameters; and estimating, based on the ellipse description of the at least one animal, a mass value of the at least one animal. The method of claim 1, wherein the neural network architecture is a segmentation network configured to: predict, on a pixel-wise basis, whether an animal is present in the input video frame; when the animal is predicted to be present in the input video frame, output a segmentation mask, wherein the segmentation mask is based upon the pixel-wise prediction, and wherein the segmentation mask represents the animal in the input video frame; and fit the segmentation mask to an ellipse, wherein the ellipse comprises values characterizing the predetermined ellipse parameters.

3. The method of claim 2, wherein the predetermined ellipse parameters comprise at least one of eccentricity, aspect ratio, and elongation.

4. The method of claim 3, wherein the ellipse description comprises eccentric area, wherein eccentric area is based upon an environment-corrected area and eccentricity.

5. The method of claim 4, wherein the environment-corrected area is calculated based upon the segmentation mask and a measurement of an environment identified in the video data.

6. The method of claim 1, wherein estimating the mass value of the animal further comprises using at least one linear regression model and at least one multiple linear regression model, both trained upon video data and animal-specific covariates.

7. The method of claim 6, wherein the animal-specific covariates comprise at least one of age, sex, and genetic information.

8. The method of claim 1, wherein the at least one animal is a rodent.

9. A system for determining animal mass, comprising: a data storage device maintaining video data representing observation of at least one animal; a processor configured to receive the video data from the data storage device and implement a neural network architecture configured to perform operations comprising: receiving an input video frame extracted from the video data; generating an ellipse description of the at least one animal based upon the input video frame, wherein the ellipse description is defined by predetermined ellipse parameters; andestimating, based on the ellipse description of the at least one animal, a mass value of the animal.

10. The system of claim 9, wherein the neural network architecture is a segmentation network configured to: predict, on a pixel-wise basis, whether an animal is present in the input video frame; when the animal is predicted to be present in the input video frame, output a segmentation mask, wherein the segmentation mask is based upon the pixel-wise prediction, and wherein the segmentation mask represents the animal in the input video frame; and fit the segmentation mask to an ellipse, wherein the ellipse comprises values characterizing the predetermined ellipse parameters.

11. The system of claim 10, wherein the predetermined ellipse parameters comprise at least one of eccentricity, aspect ratio, and elongation.

12. The system of claim 11, wherein the ellipse description comprises eccentric area, wherein eccentric area is based upon an environment-corrected area and eccentricity.

13. The system of claim 12, wherein the environment-corrected area is calculated based upon the segmentation mask and a measurement of an environment identified in the video data.

14. The system of claim 9, wherein estimating the mass value of the animal further comprises using at least one linear regression model and at least one multiple linear regression model, both trained upon video data and animal-specific covariates.

15. The system of claim 14, wherein the animal-specific covariates comprise at least one of age, sex, and genetic information.

16. A non-transitory machine-readable medium storing instructions, which when executed by at least one processor of at least one computing system, causes the system to perform operations comprising: receiving an input video frame extracted from video data representing observation of at least one animal; generating an ellipse description of the at least one animal based upon the input video frame, wherein the ellipse description is defined by predetermined ellipse parameters; and estimating, based on the ellipse description of the at least one animal, a mass value of the at least one animal.

17. The non-transitory machine-readable medium of claim 16, wherein the at least one processor of the at least one computing system is configured to implement a neural network architecture, and wherein the neural network architecture is a segmentation network configured to: predict, on a pixel-wise basis, whether an animal is present in the input video frame; when the animal is predicted to be present in the input video frame, output a segmentation mask, wherein the segmentation mask is based upon the pixel-wise prediction, and wherein the segmentation mask represents the animal in the input video frame; and fit the segmentation mask to an ellipse, wherein the ellipse comprises values characterizing the predetermined ellipse parameters.

18. The non-transitory machine-readable medium of claim 17, wherein the predetermined ellipse parameters comprise at least one of eccentricity, aspect ratio, and elongation.

19. The non-transitory machine-readable medium of claim 18, wherein the ellipse description comprises eccentric area, wherein eccentric area is based upon an environment- corrected area and eccentricity.

20. The non-transitory machine-readable medium of claim 19, wherein the environment- corrected area is calculated based upon the segmentation mask and a measurement of an environment identified in the video data.

Images