AUTOMATED B EHAVIORAL A NALYSIS U SING I NSTANCE

S EGMENTATION

A P REPRINT

Chen Yang Jeremy Forest Matthew Einhorn

Cornell University Cornell University Cornell University

Thomas A. Cleland
Department of Psychology
Cornell University
Ithaca, NY 14853
tac29@[Link]

December 14, 2023

A BSTRACT
Animal behavior analysis plays a crucial role in various fields, such as life science and biomedical
research. However, the scarcity of available data and the high cost associated with obtaining a large
number of labeled datasets pose significant challenges. In this research, we propose a novel approach
that leverages instance segmentation-based transfer learning to address these issues. By capitalizing
on fine-tuning the classification head of the instance segmentation network, we enable the tracking
of multiple animals and facilitate behavior analysis in laboratory-recorded videos. To demonstrate
the effectiveness of our method, we conducted a series of experiments, revealing that our approach
achieves exceptional performance levels, comparable to human capabilities, across a diverse range
of animal behavior analysis tasks. Moreover, we emphasize the practicality of our solution, as it
requires only a small number of labeled images for training. To facilitate the adoption and further
development of our method, we have developed an open-source implementation named Annolid (An
annotation and instance segmentation-based multiple animal tracking and behavior analysis package).
The codebase is publicly available on GitHub at [Link] This resource
serves as a valuable asset for researchers and practitioners interested in advancing animal behavior
analysis through state-of-the-art techniques.

Keywords Instance Segmentation, Multiple Object Tracking, Transfer Learning, Behavior Analysis, COCO dataset

1 Introduction
The primary objective of animal behavior analysis is to identify every instance of a specific behavior within a video and
determine its precise spatial and temporal localization. This analytical process finds practical applications in various
fields of biomedical research and notably contributes to a better understanding of the neural mechanisms involved in
both healthy and diseased behaviors. Due to the diverse array of laboratory environments and behavioral apparatuses
available (e.g., open field arenas, Morris water mazes, three-chamber assays), accurately quantifying subtle variations
in behavior, such as grooming, can pose significant challenges. Moreover, manually observing videos to extract metrics
related to the behavior of interest not only demands extensive amounts of researchers’ time but also fails to accommodate
the scale of recorded behavioral studies. This manual approach also introduces a high degree of subjectivity to the
obtained results. Automated animal pose estimation and tracking have witnessed remarkable advancements in recent
 A PREPRINT - D ECEMBER 14, 2023

years, thanks to the rapid progress in deep learning approaches [1, 2, 3, 4]. These developments have been instrumental
in revolutionizing the field. The fundamental idea revolves around training deep learning networks using annotated
keypoint data that effectively represent animals. Armed with this training, the network exhibits exceptional proficiency
in accurately estimating animal poses when presented with new video footage. The field of animal behavior analysis
encompasses a wide range of applications [5, 1, 6, 2, 3, 7, 4]. For instance, [8] developed DeepEthogram, an approach
that utilizes convolutional neural network models to process videos and extract motion features from individual frames.
The resulting models demonstrate remarkable accuracy of over 90% in classifying user-defined behaviors within fly and
mouse videos, even on a single-frame basis. The existing pose estimation techniques have a noteworthy drawback: they
primarily rely on key point representation, which lacks the flexibility to directly model animal behaviors. Additionally,
the classification at the frame level often lacks the necessary precision to accurately locate animals and their behaviors
in space. The demand for experiments that can measure multiple animals and their interactions with the environment,
including specific objects within it, is increasing. Current pose estimation-based methods do not address certain tasks
such as tracking and scoring urine deposition based on thermal camera images. Existing behavioral tracking software
often lacks or provides inaccurate tracking of multiple objects or animals. Over the past few years, significant strides
have been made in enhancing the accuracy and effectiveness of object detection and instance segmentation. These
notable advancements can be largely attributed to the remarkable progress in deep learning techniques [9, 10, 11, 12].
When it comes to tasks involving the assessment of an animal’s interaction with objects or the measurement of its
investigation area or perimeter, instance segmentation methods have emerged as a more appropriate and well-suited
approach.
Multi-Object Tracking (MOT) is a computer vision task that estimates trajectories for objects of interest in videos
[13, 14]. Many Multi-Object Tracking methods have two stages: 1) object detection and 2) data association [15]. The
first stage detects objects in each frame while the second stage associates the corresponding detections across frames to
obtain tracks. Nowadays, Tracking-by-detection is dominating the multiple object tracking pipelines by object detection
followed by temporal association or re-identification (re-ID) [16]. Zhang et al. (2020) shows that the re-ID accuracy
is mainly depending on the detection task and is not fairly learned which causes many identity switches for multiple
object tracking. Most of the appearance features are not robust to object deformations, occlusions, and illumination
variations and they also failed to discriminate objects with high similarity. The motion models use linear motion models
with constant velocity assumptions or non-linearity motion models to predict the object locations in future frames.
However, these motion models failed to deal with occlusions and interactions well and the models that tried to combine
motion and appearance modeling are having a hard time to archive a balance in the real world. In the behavior analysis
domain, the body part’s motion is the output of the neural electrical activity. Annotating and analyzing animal behaviors
from lengthy recorded videos are time-consuming tasks for scientists. With the advance of recent machine learning
methods, many toolkits have been developed to automate behavior analysis tasks to track animal centroids and recognize
user-defined behaviors, such as rearing, grooming, or fighting [8]. Bohnslav et al. (2020) model takes each frame to
predict whether each behavior is present or absent without the instance-level information. This approach does not know
where the behaviors occurred and will not work if there are multiple animals performing different behaviors. Animal
behavior analysis has benefited from the application of Multiple Object Tracking (MOT) techniques, which enable the
tracking and analysis of individual animals within a group. However, there are several challenges and issues that must
be considered when using MOT for animal behavior analysis. One challenge is ambiguity. Animals within a group
often have similar appearances, making it difficult for trackers to distinguish between them accurately. This ambiguity
can result in incorrect tracking and identity switches, compromising the reliability of the data. Occlusions pose another
challenge. Animals may obstruct each other’s movements, leading to incomplete or inaccurate tracking. This becomes
particularly problematic when animals move unpredictably or when they are in close proximity to one another. Scale
and perspective present additional hurdles. As animals move farther away from the camera, their size and shape may
change, making it challenging for trackers to maintain accurate identification. Lighting conditions can also impact
tracking accuracy. Changes in lighting can affect the visibility of animals and make it difficult for trackers to distinguish
between them accurately. Interactions and group dynamics further complicate the tracking process. Animals’ behaviors
and movements within a group are often complex and dynamic, making it challenging to identify and track individual
animals. This complexity can result in incomplete or inaccurate tracking data. Additionally, the manual annotation of
animal behavior data is time-consuming and requires significant expertise. This limitation hampers the scalability and
reproducibility of animal behavior studies. In summary, while MOT has the potential to offer valuable insights into
animal behavior, the challenges mentioned above need to be carefully considered and addressed to ensure accurate
tracking data. Such tasks are not being addressed by current pose estimation software packages like DeepLabCut
[17, 18]. Occlusions and crowded scenes also remain highly challenging and additional steps are needed to use the
body positions to classify behaviors.
To address these challenges, a new approach called annolid has been introduced. Annolid is an annotation and
segmentation-based multiple animal tracking and behavior analysis system. It aims to overcome the difficulties
associated with tracking multiple animals by providing improved annotation and segmentation tools for accurate

2
 A PREPRINT - D ECEMBER 14, 2023

tracking and behavior analysis. Annolid can help mitigate these challenges and enhance the accuracy and efficiency of
animal behavior [Link] following are the major contributions of this study:

• Demonstrating the efficacy of utilizing detailed object representations, such as segmentation, to model
behavioral patterns.
• Introducing a unique identification system where each animal is labeled with an individual ID, which is used
as the classification label for the instance segmentation network. This eliminates the need for an association
step in multiple animal tracking.
• Constructing a novel dataset specifically designed for multi-animal tracking.
• Creating and making available the annolid package, an open-source toolset for annotating, training, and
inferring segmentation networks.

2 Methods
We addressed the challenges by employing deep learning-based methods of instance segmentation, as demonstrated to
be successful in computer vision tasks [19]. Behavioral laboratory experiments, being less complex than real-world
environments, allowed us to utilize a small number of labeled images per animal type and experimental settings,
achieving state-of-the-art behavioral analysis. By leveraging pre-trained models from large datasets like ImageNet
and COCO [20], we employed transfer learning for animal behavior analysis with a limited number of labeled image
[Link] method utilizes the pre-trained Mask R-CNN model from Detectron2, which can be effectively trained
on a custom dataset of approximately 200 labeled frames in just 3 hours using Colab or within 30 minutes on an Nvidia
1080Ti GPU. Additionally, we utilized GPU-based YOLACT models, which offer real-time prediction speeds of around
30 frames per second. These models enable us to track animals by segmenting the desired region or defined behavior
and take advantage of over-fitting to achieve human-level performance with around 100 labeled frames. We evaluated
the performance of our approach on various datasets containing freely moving animals, including mice and voles. In our
study, we employed two strategies. Firstly, we fine-tuned a pre-trained instance segmentation network, Mask-RCNN
[9], to achieve highly accurate predictions. Alternatively, we utilized faster instance segmentation networks called
YOLACT [12] and YOLACT++ [11], which provide real-time performance but sacrifice a bit of accuracy. These
models simultaneously segmented and tracked animals in each video frame. To ensure accurate instance separation
and prevent confusion between animals, we assigned a unique class label to each animal (e.g., animal_1, animal_2, ...,
animal_X). This labeling scheme enabled the network’s classification head to learn distinct appearance features for
different instances, minimizing ID switches or confusion among animals. Consequently, re-identifying animals across
frames was avoided by associating the Regions of Interest (ROIs) throughout the video sequence.
Introducing Annolid (Annotation and Annelid with segmentation), we presented a novel behavioral software that
utilizes instance segmentation for multiple object tracking and behavior analysis. Leveraging transfer learning from
pre-trained instance segmentation networks based on the COCO dataset [20], Annolid automatically segments and
tracks animals and objects while classifying user-defined behaviors. This approach significantly reduces the need for
extensive labeling, requiring only a small number of labeled frames (approximately 100). Annolid offers a unified
graphical user interface that simplifies dataset creation and visualization. It supports training instance segmentation
models using popular methods such as Mask R-CNN [19], PointRend [10] implemented in Detectron2 [21], YOLACT
[12], and YOLACT++ [11]. These models can also be used for real-time prediction on new data. However, it is
important to note that annotating and training the model for each new video featuring an unseen animal is necessary. Our
experimental results demonstrate that our approach achieves performance levels comparable to humans, even without
a large set of labeled images or the need for separate recognition networks such as Siamese networks [22] for face
recognition. This is particularly valuable because data acquisition for specific animals is often limited in laboratories,
and the cost associated with labeling a large number of training examples is high. Using Annolid is straightforward and
follows a recommended sequence of steps:

Step Description
1 Selection of desired frames for labeling.
2 Labeling the selected frames.
3 Conversion and creation of the dataset in COCO format.
4 Training the selected network.
5 Scoring the video using the trained model.
6 Performing statistical analysis on the tracking results.
Table 1: Recommended sequence of steps using Annolid.

3
 A PREPRINT - D ECEMBER 14, 2023

We have developed a user-friendly Graphical User Interface (GUI) based on the labelme [23] tool to assist users
throughout their workflow. In addition, Annolid offers users the flexibility to choose the instance segmentation method
that best suits their datasets. They can opt for YOLACT, YOLACT++, Mask-RCNN, or PointRend. Our main
contribution lies in adopting instance segmentation networks for tracking. We treat each instance as its own class frame
by frame, eliminating the need for the association step required in traditional Multiple Object Tracking (MOT) methods.
By doing so, Annolid achieves both instance segmentation and data association without the need for additional track
matching. Furthermore, we have eliminated the requirement to maintain tracklets since we have precise information
about the location and identities of objects in each frame. Figure 1 shows an overview of Annolid’s workflow.

Figure 1: Overview of Annolid workflow

2.1 Integration of the Segment Anything Model(SAM) into Annolid

In this section, we discuss the seamless integration of the Segment Anything Model, introduced by Kirillov et al. in their
2023 paper [25], into the Annolid platform. The Segment Anything model, developed and released by Meta research,
has undergone extensive training with a vast dataset comprising more than one billion masks. This training has endowed
the model with a high level of sophistication and the ability to accurately identify and delineate instances within
[Link] of the key objectives of integrating the Segment Anything model into Annolid is to simplify the instance
labeling process for users. Traditionally, instance segmentation tasks can be time-consuming and labor-intensive,
requiring meticulous manual annotation. Within the Annolid platform, users are presented with an intuitive interface
that allows them to label instances with remarkable ease. Users need only designate a few positive and negative points
within the image. These points serve as guidance for the Segment Anything model. It automatically generates precise
masks, which are then processed into polygons outlining the identified instances within the image. These polygons can
be further simplified using map generalization techniques commonly employed in geographic information systems
(GIS), as implemented in [26]. This automation significantly reduces the time and effort typically associated with
manual polygon drawing and instance delineation. While the automatic polygon generation is highly accurate, Annolid
provides users with the flexibility to fine-tune the generated polygons. This feature ensures that users can achieve the
level of precision required for their specific labeling tasks. Beyond instance delineation, users have the capability to
assign class labels to the generated polygons based on their research requirements, even assigning behaviors directly
as the polygon name. The integration of the SAM into Annolid promises a significant improvement in efficiency and
accuracy in the labeling process.

2.2 Classification head

To distinguish between instances in a given video, we utilized the classification head of the instance segmentation
model. Specifically, we labeled each instance as a distinct class (e.g., "mouse1" and "mouse2") across frames, instead
of assigning a single class label such as "mouse". The unique identity assigned to each instance was then used as the

4
 A PREPRINT - D ECEMBER 14, 2023

Figure 2: Overview of the Annolid GUI based on labelme[23].

Figure 3: Example frames of mice with manually annotated masks.

training class label in the instance segmentation’s classification branch. This approach ensures that the network is
trained to classify instances correctly, even in frames that were not labeled. However, due to motion and occlusion, the
same instance can disappear in some frames either fully or partially (e.g., due to body part occlusion). To account for
this, the model should be given a set of frames representing different appearances of the same object throughout the
video. Various methods are available for extracting frames for labeling, such as random sampling, optical flow, and
keyframes. In our experiments, we used a random sampling method for selecting training frames to label, as it was
faster.

5
 A PREPRINT - D ECEMBER 14, 2023

Figure 4: Human in the loop with model auto-labeling.

Figure 5: Example: Predicted polygons by a trained instance segmentation model. In this use case, each polygon
represents either huddling or lone behavior.

Figure 6: An example of predicted body parts on the dataset[24], treating pose estimation as a special instance
segmentation case.

6
 A PREPRINT - D ECEMBER 14, 2023

2.3 Datasets and Metrics

In order to label and prepare datasets, we utilize our Annolid user interface, which is based on Labelme. This interface
allows us to create polygons for instances and regions of interest for segmentation masks. These masks can be used for
various purposes, such as identifying behavioral interactions or body parts. The interface generates Labelme ".json" files,
which can be converted to COCO format datasets. These datasets are suitable for training instance segmentation-based
models like YOLACT or Mask-RCNN. When converting to COCO format, both a training set and a validation set
are created. For cases involving multiple animal pose estimation, the same user interface can be used to label the key
points, which can then be converted into small regions of masks. The segmentation network treats pose estimation
as a specialized use case. There are no limitations on what a user can label. In the most extreme case, every pixel
in the video frame could be labeled. Our approach, referred to as Annolid, consists of four stages: preprocessing,
labeling, training, motion tracking, and behavior analysis. To evaluate the efficacy of Annolid for segmenting instances,
we meticulously labeled 200 frames randomly selected from a video featuring three interacting mice. The datasets
provide annotations for combined instance masks and identity information, represented as class plus ID numbers
(e.g., LeftMouse, RightMouse, and Mouse), as illustrated in Figure 3. Our trained models exhibit robust performance
across diverse scenarios, effectively handling challenges such as multiple animals in rapid motion, occlusion at varying
locations, and temporal variations from the camera. Notably, the model achieved an impressive average precision (AP)
of 44.65 using 80 labeled frames in the training set. Furthermore, we extended our efforts to curate a Multi-Animal
Tracking Dataset (MATD) 2023. This dataset comprises ten individual videos, each presenting challenging scenarios
involving fast motion, occlusion, and other complex visual factors. Notably, the average length of each video in MATD
2023 exceeds 1000 frames, surpassing the maximum video duration observed in previous Video Object Segmentation
(VOS) benchmarks. The datasets are available upon request.

2.4 Training Instance Segmentation Models

We applied a standard transfer learning strategy using the pre-trained weights from a network trained on the COCO
datasets [20].We fine-tuned the network on custom-labeled datasets. The YOLACT models were trained on a computer
with an NVIDIA 1080Ti GPU card. The operating system was Windows 10 with CUDA 10.0, CUDNN 7.0, and
PyTorch 1.4.0. With a single GPU on this machine, it takes a few hours for several thousand iterations. Inferencing
on new videos is about 30 FPS in real-time on the trained model on the same GPU. We also provided Google Colab
notebooks for training models. In addition to the strong performance, the approach is very simple and runs at 30 FPS
on a single NVIDIA GTX 1080Ti GPU. It helps the scientists to adopt developing software by showing examples with
neural networks. The input image is resized to 512 x 512, and the training step takes about 2 hours on one GTX 1080 Ti
GPU. For training Mask-RCNN in Google Colab, we used Resnet50 as our default backbone with the model parameters
pre-trained on the ImageNet [27] and COCO dataset [20] were used to initialize our model. We trained our model with
the default optimizer implemented in Detectron2 for 3000 epochs with a starting learning rate of 0.00025. The batch
size is set to 8. Other parameters were unchanged as the default configurations for Mask-RCNN FPN 3x as described in
Detectron2 [21].

3 Inferences
During inference, we performed threshold-based filtering on the class scores predicted by the network. We did not need
to follow the standard tracking methods to associate across frames. In our case, in each frame, the model was trained to
predict the locations and identities of the objects.

3.1 Evaluation

In the evaluation of tracking accuracy for the test set within the two voles dataset, we employed the Multiple Object
Tracking Accuracy (MOTA) metric [28]. The dataset encompasses a total of 10,575 frames within the video. These
results stem from predictions made by a model trained on a subset of 368 labeled frames.
The MOTA metric is defined as follows:
F alseN egatives + IDSwitches + F alseP ositives
M OT A = (1 − ) (1)
NGT
.
In our evaluation, we observed 52 instances of ID switches, 38 False Negative frames, and 11 False Positive frames.
Applying the MOTA formula, the calculated MOTA score for this test video is 99.04%. We conducted an experiment
using varying numbers of labeled frames in the training set, as depicted in Figure 10. To evaluate the performance of
instance segmentation, we employed Average Precision (AP).

7
 A PREPRINT - D ECEMBER 14, 2023

Figure 7: AP increases by training with more labeled frames

3.2 Results

In this section, we present the results on the Multi-Animal Tracking Dataset (MATD) 2023 in table 2. Since track-
ing based on instance segmentation can be seen as the classification of instance masks, we directly apply the AP
measurements to evaluate the trained instance model.

Table 2: Results on instance segmentation using the multi-animal tracking validation dataset.
AP ↑ AP50 ↑ AP75 ↑ APS ↑ APM ↑ APL ↑
birds 64.735 86.246 75.953 - 39.360 65.971
fish 65.806 96.153 82.054 - 58.408 73.087
mice 79.906 97.006 93.182 - 79.906 0.0
frogs 30.573 66.584 21.457 27.287 42.167 -
voles 51.711 94.442 43.285 49.414 62.465 -
polar bears 66.146 73.102 73.102 - 84.950 67.426
ants 36.121 92.616 12.404 44.455 36.429 -
bees 69.530 87.624 87.624 - - 69.530
white mice 69.328 71.926 71.926 - - 69.328
gerbils 66.349 90.064 71.211 - 68.735 68.298

In this study, we present qualitative results obtained from the two voles dataset, visualized in Figure 9. Our proposed
approach exhibits remarkable performance, notably reducing the occurrence of ID switches. Moreover, our method
achieves inference speeds approaching real-time, especially noteworthy for YOLACT models. These results are
compelling, given the simplicity of our approach, which refrains from introducing complexity for frame-to-frame
association. Furthermore, we conducted a comprehensive evaluation of Annolid’s performance across varying amounts
of labeled training data. Impressively, utilizing approximately 100 labeled frames for training yielded high Multiple
Object Tracking Accuracy (MOTA), approaching human-level performance. In a test video featuring two voles and
comprising a total of 10,575 frames, we computed 52 ID switches (0.49% of frames), 38 false negatives (0.36%),
and 11 false positives (0.1%). The resulting MOTA score reached an impressive 99.05%. Figure 9 showcases 40
randomly selected tracking frames, with each row presenting sampled frames in chronological order. The images are

8
 A PREPRINT - D ECEMBER 14, 2023

Figure 8: Overview of tracks for four voles in a 6-minute stitched video.

annotated with segmentation masks and identities, with distinct colors representing individual animals. For optimal
visual appreciation, we recommend viewing the figure in color.

3.3 Qualitative Results

In this section, we present a selection of qualitative results from the tracking process. For additional use cases and more
examples, please refer to the appendix. Figure 6 showcases exemplary results of our pose estimation, highlighting
precise tracking of specific body parts such as the nose, left ear, right ear, body centroid, and tail base. Annolid
demonstrates successful application across diverse experimental videos, showcasing its capability to track segmented
regions of interest according to users’ specific needs and [Link] on the results shown in Figure 10 and
Figure 11, our method successfully demonstrates instance segmentation and recognition even in complex scenarios
such as two animals crossing over each other, while effectively recovering from occlusions.

4 Discussion
In our implementation, we embrace a multiple-object tracking classification approach that extracts information solely
from the current frames, distinct from batch methods that rely on future frames. Online methodologies, exemplified by
SORT [29] and DeepSort [30], address the data association challenge by deploying techniques such as the Kalman
Filter [31] for object motion prediction and the Hungarian algorithm [32] for data association. Notably, our work
departs from utilizing motion information derived from adjacent frames to enhance tracking accuracy. In contrast, Hou
et al. [33] introduced a Tube Convolutional Neural Network (T-CNN) in a study, leveraging 3D convolution features in
equal-length video clips for action recognition and localization. While we don’t incorporate motion information from
adjacent frames in our current work, our future exploration involves integrating networks like T-CNN. This investigation
aims to uncover potential advantages, particularly in the realm of self-supervised learning. In our envisioned future work,
we plan to delve into the integration of networks such as T-CNN and explore the potential benefits of self-supervised
learning. This approach holds promise for reducing annotation efforts, rendering Annolid more attractive for diverse

9
 A PREPRINT - D ECEMBER 14, 2023

Figure 9: Randomly sampled tracking results of Annolid for two interacting voles, with color-coded masks and bounding
boxes. For better visibility, zoom in to see the colors in detail.
10
 A PREPRINT - D ECEMBER 14, 2023

Figure 10: An example of detected urine spots, displayed with bounding boxes.

applications, especially in scenarios where behavior is influenced by disease or genetic engineering. The use of
self-supervised learning enables investigators to circumvent bias introduced through manual labeling, allowing the
model to learn behavior directly through instance segmentation.

5 Conclusion
In this paper, we present our instance segmentation approach, which effectively improves tracking accuracy without
requiring extensive engineering efforts. The results demonstrate that our approach offers a significant advantage in
behavioral experiments by not being heavily reliant on labeled data. We introduce Annolid, a framework that achieves
comparable performance to pose estimation and multiple object tracking (MOT) methods on our benchmark datasets,
in terms of both tracking accuracy and inference speed. Furthermore, Annolid inherently exhibits data-efficiency by
leveraging network overfitting capabilities, making it highly valuable for applications involving multiple animal tracking
and behavior analysis.

6 Acknowledgements
This work was supported by NSF grant 1743214.

11
 A PREPRINT - D ECEMBER 14, 2023

References
[1] Mathis, A., Schneider, S., Lauer, J. & Mathis, M. W. A primer on motion capture with deep learning: Principles,
pitfalls, and perspectives. Neuron 108, 44–65 (2020).
[2] Pereira, T. et al. Fast animal pose estimation using deep neural networks. bioRxiv (2018).
URL [Link] [Link]
content/early/2018/05/30/[Link].
[3] Pereira, T. D. et al. Sleap: Multi-animal pose tracking. bioRxiv (2020). URL [Link]
org/content/early/2020/09/02/2020.08.31.276246. [Link]
2020/09/02/[Link].
[4] Lauer, J. et al. Multi-animal pose estimation and tracking with deeplabcut. bioRxiv (2021).
[5] Arac, A., Zhao, P., Dobkin, B. H., Carmichael, S. T. & Golshani, P. Deepbehavior: A deep learning toolbox for
automated analysis of animal and human behavior imaging data. Frontiers in Systems Neuroscience 2019 May
7;13:20 (2019).
[6] Segalin, C. et al. The mouse action recognition system (mars): a software pipeline for automated analysis of
social behaviors in mice. bioRxiv (2020).
[7] Sun, J. J. et al. The multi-agent behavior dataset: Mouse dyadic social interactions. arXiv preprint
arXiv:2104.02710 (2021).
[8] Bohnslav, J. P. et al. Deepethogram: a machine learning pipeline for supervised behavior classification from
raw pixels. bioRxiv (2020). URL [Link]
312504. [Link]
[9] He, K., Gkioxari, G., Dollár, P. & Girshick, R. Mask r-cnn. In Proceedings of the IEEE international conference
on computer vision, 2961–2969 (2017).
[10] Kirillov, A., Wu, Y., He, K. & Girshick, R. PointRend: Image segmentation as rendering (2019).
[11] Bolya, D., Zhou, C., Xiao, F. & Lee, Y. J. Yolact++: Better real-time instance segmentation (2019). 1912.06218.
[12] Bolya, D., Zhou, C., Xiao, F. & Lee, Y. J. Yolact: Real-time instance segmentation. In ICCV (2019).
[13] Zhou, X., Koltun, V. & Krähenbühl, P. Tracking objects as points. arXiv preprint arXiv:2004.01177 (2020).
[14] Sun, S., Akhtar, N., Song, H., Mian, A. & Shah, M. Deep affinity network for multiple object tracking (2018).
1810.11780.
[15] Ciaparrone, G. et al. Deep learning in video multi-object tracking: A survey. Neurocomputing 381, 61–88 (2020).
[16] Zhang, Y., Wang, C., Wang, X., Zeng, W. & Liu, W. Fairmot: On the fairness of detection and re-identification in
multiple object tracking (2020). 2004.01888.
[17] Mathis, A. et al. Deeplabcut: markerless pose estimation of user-defined body parts with deep learning. Nature
Neuroscience (2018). URL [Link]
[18] Mathis, M. W. & Mathis, A. Deep learning tools for the measurement of animal behavior in neuroscience. Current
Opinion in Neurobiology 60, 1–11 (2020). URL [Link]
[19] He, K., Gkioxari, G., Dollar, P. & Girshick, R. Mask r-cnn. In Proceedings of the IEEE International Conference
on Computer Vision (ICCV) (2017).
[20] Lin, T.-Y. et al. Microsoft coco: Common objects in context. In European conference on computer vision,
740–755 (Springer, 2014).
[21] Wu, Y., Kirillov, A., Massa, F., Lo, W.-Y. & Girshick, R. Detectron2. [Link]
facebookresearch/detectron2 (2019).
[22] Koch, G. Siamese neural networks for one-shot image recognition (2015).
[23] Wada, K. labelme: Image Polygonal Annotation with Python. [Link]
(2016).
[24] Nath, T. et al. Using deeplabcut for 3d markerless pose estimation across species and behaviors. Nature protocols
14, 2152–2176 (2019).
[25] Kirillov, A. et al. Segment anything. arXiv:2304.02643 (2023).
[26] Hügel, S. Simplification (2021). URL [Link]

12
 A PREPRINT - D ECEMBER 14, 2023

[27] Deng, J. et al. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision
and pattern recognition, 248–255 (Ieee, 2009).
[28] Bernardin, K., Elbs, A. & Stiefelhagen, R. Multiple object tracking performance metrics and evaluation in a smart
room environment. In Sixth IEEE International Workshop on Visual Surveillance, in conjunction with ECCV,
vol. 90, 91 (Citeseer, 2006).
[29] Wojke, N., Bewley, A. & Paulus, D. Simple online and realtime tracking with a deep association metric. In 2017
IEEE International Conference on Image Processing (ICIP), 3645–3649 (IEEE, 2017).
[30] Wojke, N. & Bewley, A. Deep cosine metric learning for person re-identification. In 2018 IEEE Winter Conference
on Applications of Computer Vision (WACV), 748–756 (IEEE, 2018).
[31] Welch, G., Bishop, G. et al. An introduction to the kalman filter (1995).
[32] Kuhn, H. W. The hungarian method for the assignment problem. Naval research logistics quarterly 2, 83–97
(1955).
[33] Hou, R., Chen, C. & Shah, M. Tube convolutional neural network (t-cnn) for action detection in videos (2017).
1703.10664.

13
 A PREPRINT - D ECEMBER 14, 2023

A Appendix
A.1 Detect deposited urine

Figure 11: An example of detected newly deposited urine display, highlighted with bounding boxes.

A.2 Tracking results of Annolid for three interacting voles and their body parts

14
 A PREPRINT - D ECEMBER 14, 2023

Figure 12: Randomly sampled tracking results of Annolid for three interacting voles and their body parts, with color-
coded masks. For optimal viewing, zoom in to see the colors
15 and details.