SqueakPose Studio: An end-to-end platform for pose estimation and real-time edge-AI deployment
Curation statements for this article:-
Curated by eLife
eLife Assessment
This important work introduces an integrated open-source platform for behavioral acquisition and pose estimation that substantially improves the accessibility and speed of real-time animal tracking workflows. The evidence supporting the utility and usability of SqueakPose Studio is compelling, particularly the substantial inference speed gains, intuitive graphical interface, flexible pose configuration, and successful testing on independent datasets, although the evidence supporting broader benchmarking claims and the hardware ecosystem surrounding MouseHouse and SqueakView remains somewhat incomplete. The study will be of broad interest to neuroscientists and behavioural researchers seeking scalable and user-friendly approaches for real-time behavioral analysis, and the work would be further strengthened by more rigorous benchmarking, expanded installation and hardware documentation, formal software release practices, and clearer delineation between demonstrated capabilities and future applications.
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
Accurate pose estimation underpins quantitative analysis of behavior, yet many deep learning-based tracking tools remain optimized for offline workflows that rely on fragmented software pipelines, workstation-grade GPUs, or external middleware to enable real-time deployment. Here we present an integrated software-hardware ecosystem for pose estimation that spans dataset creation, model training, offline analysis, and real-time deployment on embedded edge-computing devices. SqueakPose Studio provides a software suite for whole-frame, deep learning-based pose estimation that unifies dataset creation, manual and model-assisted labeling, model training, validation, and large-scale offline inference. The system leverages modern object-detection architectures to enable efficient end-to-end training and inference without patch-based sampling or multi-stage postprocessing, and supports execution on CPUs, GPUs, and Apple Silicon. For experimental settings requiring continuous recording and synchronized data acquisition, SqueakView enables real-time model deployment, video capture, and sensor logging on embedded edge-computing hardware, while MouseHouse provides a compact, modular enclosure designed for home cage-based experiments that integrates embedded GPU compute, microcontroller-based timing, and peripheral I/O. A shared data format and deterministic timing architecture ensure consistency across offline analysis and real-time deployment. Together, SqueakPose Studio, SqueakView, and MouseHouse provide a unified platform for pose estimation that supports both conventional offline analysis and embedded, real-time experimentation, without reliance on workstation-grade hardware or external middleware.
Article activity feed
-
eLife Assessment
This important work introduces an integrated open-source platform for behavioral acquisition and pose estimation that substantially improves the accessibility and speed of real-time animal tracking workflows. The evidence supporting the utility and usability of SqueakPose Studio is compelling, particularly the substantial inference speed gains, intuitive graphical interface, flexible pose configuration, and successful testing on independent datasets, although the evidence supporting broader benchmarking claims and the hardware ecosystem surrounding MouseHouse and SqueakView remains somewhat incomplete. The study will be of broad interest to neuroscientists and behavioural researchers seeking scalable and user-friendly approaches for real-time behavioral analysis, and the work would be further strengthened by more …
eLife Assessment
This important work introduces an integrated open-source platform for behavioral acquisition and pose estimation that substantially improves the accessibility and speed of real-time animal tracking workflows. The evidence supporting the utility and usability of SqueakPose Studio is compelling, particularly the substantial inference speed gains, intuitive graphical interface, flexible pose configuration, and successful testing on independent datasets, although the evidence supporting broader benchmarking claims and the hardware ecosystem surrounding MouseHouse and SqueakView remains somewhat incomplete. The study will be of broad interest to neuroscientists and behavioural researchers seeking scalable and user-friendly approaches for real-time behavioral analysis, and the work would be further strengthened by more rigorous benchmarking, expanded installation and hardware documentation, formal software release practices, and clearer delineation between demonstrated capabilities and future applications.
-
Reviewer #1 (Public review):
This is a well-written and fully documented methods paper.
The authors have established a clear rationale for their new packages, especially for real-time use, and demonstrate significant speed improvements that will likely appeal to many users of tools like DLC, SLEAP, and LightningPose. The inclusion of a graphical user interface will help make the package more accessible to neuroscientists with limited computational expertise. While it may be challenging to get users to switch from their established workflows for video analysis, the speed gains offered by this package make it worth considering. The hardware aspects of the project are well-documented, and the GitHub repository for this part of the setup is also thorough. Overall, this paper provides a clear summary of the tools, their uses, setup, and …
Reviewer #1 (Public review):
This is a well-written and fully documented methods paper.
The authors have established a clear rationale for their new packages, especially for real-time use, and demonstrate significant speed improvements that will likely appeal to many users of tools like DLC, SLEAP, and LightningPose. The inclusion of a graphical user interface will help make the package more accessible to neuroscientists with limited computational expertise. While it may be challenging to get users to switch from their established workflows for video analysis, the speed gains offered by this package make it worth considering. The hardware aspects of the project are well-documented, and the GitHub repository for this part of the setup is also thorough. Overall, this paper provides a clear summary of the tools, their uses, setup, and benefits.
I have a few minor questions about the collective set of tools.
First, the GitHub repository for SqueakPoseStudio appears to be missing a testing routine and associated badge, and the package has not been formally released. This means users would need to download the repository to install it, correct? I suggest the authors consider publishing a formal release of the package, making it installable via pip, and including a basic testing routine to clearly display the package's status on the repository page. Adding a DOI from Zenodo would also be helpful. A testing routine is especially useful when updates are made, as many users avoid repositories with failing tests.
Second, the installation instructions simply state "Create a virtualenv and install:". This may not be sufficient for many researchers, as most neuroscientists are not experienced Python programmers and require clear guidance on the environment specific to this package. The installation instructions should be expanded to provide more detailed guidance and encourage more users. It would also be helpful to verify that the setups work across Windows, Mac, and Linux.
Third, the package defaults to UMAP for non-linear dimensionality reduction, which has some known issues. Can the package be modified to allow for alternative mapping methods, such as PaCMAP, PyDiffMap, or the more comprehensive topometry package?
Finally, what specific GPUs have been tested with the package, and are there any limitations based on the age of the video card or the available libraries for the deep learning component of the package?
-
Reviewer #2 (Public review):
Summary:
This work presents three tools: SqueakPose Studio, which is used for pose estimation; SqueakView, which is used for real-time video and sensor data capture and analysis; and MouseHouse, which is a behavioral and sensor suite for mouse experiments. Together, these tools provide a comprehensive behavioral platform for acquiring and analyzing video, sensor, and behavioral data. The work is open source and provided as a resource for the field.
Strengths:
(1) Squeakpose Studio was relatively easy to install and use. We were impressed that we were able to install it and test our own videos with minimal struggles. The authors provide installation tutorial videos that were very helpful.
(2) The GUI environment for SqueakPose Studio was very usable, and the authors should be commended on the time and effort …
Reviewer #2 (Public review):
Summary:
This work presents three tools: SqueakPose Studio, which is used for pose estimation; SqueakView, which is used for real-time video and sensor data capture and analysis; and MouseHouse, which is a behavioral and sensor suite for mouse experiments. Together, these tools provide a comprehensive behavioral platform for acquiring and analyzing video, sensor, and behavioral data. The work is open source and provided as a resource for the field.
Strengths:
(1) Squeakpose Studio was relatively easy to install and use. We were impressed that we were able to install it and test our own videos with minimal struggles. The authors provide installation tutorial videos that were very helpful.
(2) The GUI environment for SqueakPose Studio was very usable, and the authors should be commended on the time and effort that went into improving the useability of their system. The keypoint and skeleton configuration was flexible, allowing us to define custom body part sets without modifying code directly. The pose estimation accuracy on our own videos was good right out of the box, without requiring fine-tuning or retraining. For a tool being evaluated for the first time, this was all very impressive!
Weaknesses:
(1) While we were able to install and test Squeakpose Studio, it was not entirely seamless. The primary installation resource is a tutorial video, and we would recommend supplementing this with a written installation checklist that explicitly lists all required software dependencies (e.g. Python, UV, Visual Studio). The tutorial video was also at times unclear in distinguishing required from optional components. For example, Visual Studio is described as not necessary, yet the tutorial demonstrates the workflow entirely within that environment, so it may be challenging for a user to follow along without that. We recommend that the authors adopt a stricter, step-by-step installation guide that is prescriptive about required software and leaves little room for confusion.
(2) The paper also describes SqueakView and MouseHouse. Unfortunately, we were unable to evaluate these components as both require the MouseHouse hardware platform. Even without directly using MouseHouse, we noticed some incompleteness here, as we could not locate a bill of materials, component pricing, or assembly guide in the paper or associated GitHub repositories. Given that affordability and accessibility are central claims, a consolidated parts list, approximate costs, and a build guide or video would be necessary for most labs to realistically decide whether they plan to replicate the hardware and evaluate this functionality that the paper describes. In this regard, we felt that MouseHouse and potentially SqueakView were not sufficiently documented for publication.
(3) The benchmarking comparison to DeepLabCut (DLC) introduced multiple challenges that left us unclear if the head-to-head comparison was appropriate as described. First, the dataset used for benchmarking was small and homogeneous, from the methods they used "10 min open-field tasks of single mice with bilateral photometry cables." As such, the claims about comparisons between SqueakPose Studio and DLC may be too broad, given this single test case. Specifically, this dataset does not test robustness across lighting conditions, coat colors, species, occlusions, different-shaped arenas, etc. Second, the comparison to DLC in Figure 1 does not include any quantitative statistical comparisons, which are needed to evaluate the claims that were made. For instance, the error in Figure 1e looks worse for their system than DLC, although statistical comparisons were not made. Third, there are many settings and optimizations that can be made for both systems. Without more detail, this makes it hard to know if the head-to-head comparison is really fair. Fourth - the metrics are given as very specific numbers from single runs, i.e., an inference time of 71.59 minutes in Figure 1d. This metric would be more meaningful if it reported the mean of multiple runs, with error estimation. Finally, while the code is available, the trained datasets are made available only on "reasonable request". Given the importance of these datasets to evaluating the method and allowing others to benchmark it against other systems, these should be made available on GitHub. Overall, I would recommend toning down the comparison to DLC and focusing on the strengths of Squeakpose Studio on its own merits.
(4) The paper at times makes general statements that are beyond what is shown. For instance, discussions of use in human applications are aspirational and should be treated much more conservatively in the discussion, or possibly even removed. As it stands, the discussion implies that this system can already do "zero-shot tracking of human posture and movement", enabling "a bridge between preclinical and clinical behavioral analysis". In principle, this may be true, but even for a Discussion section, this goes far beyond the capabilities that the paper actually shows.
(5) While the comprehensive nature of the system and its 3 parts is impressive, I felt that it also detracted from the main focus of the paper, which was Squeakpose Studio. I might recommend dropping the other two parts, as they also require a much higher bar for a user to evaluate, and only present the Squeakpose Studio in this paper, presenting this as a general resource for pose estimation. This would also allow them more space to more comprehensively benchmark SqueakPose Studio.
-
Author response:
Public Reviews:
Reviewer #1 (Public review):
This is a well-written and fully documented methods paper.
The authors have established a clear rationale for their new packages, especially for real-time use, and demonstrate significant speed improvements that will likely appeal to many users of tools like DLC, SLEAP, and LightningPose. The inclusion of a graphical user interface will help make the package more accessible to neuroscientists with limited computational expertise. While it may be challenging to get users to switch from their established workflows for video analysis, the speed gains offered by this package make it worth considering. The hardware aspects of the project are well-documented, and the GitHub repository for this part of the setup is also thorough. Overall, this paper provides a clear summary of …
Author response:
Public Reviews:
Reviewer #1 (Public review):
This is a well-written and fully documented methods paper.
The authors have established a clear rationale for their new packages, especially for real-time use, and demonstrate significant speed improvements that will likely appeal to many users of tools like DLC, SLEAP, and LightningPose. The inclusion of a graphical user interface will help make the package more accessible to neuroscientists with limited computational expertise. While it may be challenging to get users to switch from their established workflows for video analysis, the speed gains offered by this package make it worth considering. The hardware aspects of the project are well-documented, and the GitHub repository for this part of the setup is also thorough. Overall, this paper provides a clear summary of the tools, their uses, setup, and benefits.
We thank this reviewer for the positive comments and have provided responses to the specific and constructive questions listed below.
I have a few minor questions about the collective set of tools.
First, the GitHub repository for SqueakPoseStudio appears to be missing a testing routine and associated badge, and the package has not been formally released. This means users would need to download the repository to install it, correct? I suggest the authors consider publishing a formal release of the package, making it installable via pip, and including a basic testing routine to clearly display the package's status on the repository page. Adding a DOI from Zenodo would also be helpful. A testing routine is especially useful when updates are made, as many users avoid repositories with failing tests.
We thank the reviewer for this helpful suggestion. We agree that visible testing improves user confidence and reproducibility.
SqueakPose Studio is currently distributed through a repository-based uv workflow rather than through PyPI alone. This is intentional. The application depends on platform-specific deep-learning libraries, and cloning the repository followed by uv sync provides a reproducible environment across Linux, macOS, and Windows while allowing the application to select CUDA, Apple MPS, or CPU execution at runtime. The written installation instructions now clearly describe this workflow.
In response to the reviewer’s suggestion, we have added a unit-test suite covering the core helper modules used for label handling, dataset export, prediction, inference, and training logic. We have also added an automated GitHub Actions workflow that runs the tests on pushes and pull requests, together with a repository badge that displays the current test status.
Second, the installation instructions simply state "Create a virtualenv and install:". This may not be sufficient for many researchers, as most neuroscientists are not experienced Python programmers and require clear guidance on the environment specific to this package. The installation instructions should be expanded to provide more detailed guidance and encourage more users. It would also be helpful to verify that the setups work across Windows, Mac, and Linux.
We agree that installation guidance should be accessible to researchers who may not routinely manage Python environments. In addition to the existing video walkthrough, we have expanded the written GitHub documentation to provide a clearer, step-by-step installation checklist.
The revised README now distinguishes required components from optional tools, explains the repository-based uv workflow, and provides the minimal commands needed to create the managed environment and launch the application.
We have also clarified that an integrated development environment is optional. Although Visual Studio Code is used in the tutorial as a convenient interface for demonstrating the workflow, users may launch SqueakPose Studio directly from a terminal and are not required to use Visual Studio Code, Visual Studio, or any other editor.
We have tested the application on Apple Silicon macOS systems, Windows systems, Linux systems, and NVIDIA GPU-enabled machines. SqueakPose Studio selects CUDA, Apple MPS, or CPU execution at runtime according to availability. Because accelerator support is partly determined by upstream packages such as PyTorch and Ultralytics, we have added links to the relevant compatibility documentation so that users can confirm whether their current hardware and driver configuration are supported.
Third, the package defaults to UMAP for non-linear dimensionality reduction, which has some known issues. Can the package be modified to allow for alternative mapping methods, such as PaCMAP, PyDiffMap, or the more comprehensive topometry package?
We agree with the reviewer that UMAP has limitations and that no single nonlinear dimensionality-reduction method is optimal for all pose datasets or behavioral questions.
In SqueakPose Studio, the UMAP/HDBSCAN workflow is included as an accessible exploratory example for dimensionality reduction and clustering of pose-derived features. Our goal was not to designate UMAP as a preferred or definitive analysis method, but to provide an interpretable starting point that allows users to identify candidate clusters and inspect representative videos to evaluate what the embedding is capturing.
We agree that supporting additional approaches, such as PaCMAP, PyDiffMap, or related tools, could be useful, and we will consider adding these as modular options in future versions. At the same time, SqueakPose Studio is not intended to replace specialized downstream behavioral-analysis packages or to adjudicate which embedding method is best for a particular dataset. Pose outputs can be exported for downstream analysis in other environments, including CEBRA, Keypoint-MoSeq, and packages implementing alternative clustering or dimensionality-reduction approaches.
We have clarified in the documentation that the included UMAP/HDBSCAN workflow is intended as an exploratory demonstration rather than as a required or privileged analysis pipeline.
Finally, what specific GPUs have been tested with the package, and are there any limitations based on the age of the video card or the available libraries for the deep learning component of the package?
As noted above, GPU compatibility is determined by the deep-learning and hardware-acceleration libraries on which SqueakPose Studio depends, including PyTorch, Ultralytics, CUDA, Apple MPS, and ROCm. Our development ethos is to track current stable versions of these packages rather than maintain separate legacy dependency stacks. This improves performance, simplifies support, and allows users to benefit from ongoing improvements in upstream libraries, but it also means that older GPU architectures may lose support as they are deprecated by those upstream tools.
For NVIDIA systems, the current package is indexed against CUDA 13.2. CUDA 13.x has deprecated support for some older GPU architectures, so users with older NVIDIA cards may need to use CPU inference or upgrade hardware. However, CUDA 13 is supported on GeForce RTX 20-series, 30-series, 40-series, 50-series, and professional equivalents. We made this clearer in the documentation and provided links to upstream CUDA, PyTorch, and Ultralytics compatibility resources so users can determine whether their hardware is supported.
For Apple Silicon, the package can use PyTorch MPS acceleration, which supports M-series chips. For AMD GPUs, we do not currently maintain AMD-specific test hardware, but PyTorch supports ROCm on Linux for supported AMD GPUs. ROCm support is more limited on Windows, so AMD users should consult the current PyTorch ROCm compatibility documentation.
Overall, our support commitment is to maintain compatibility with current upstream deep-learning frameworks rather than to guarantee support for all older or vendor-specific GPU configurations.
Reviewer #2 (Public review):
Summary:
This work presents three tools: SqueakPose Studio, which is used for pose estimation; SqueakView, which is used for real-time video and sensor data capture and analysis; and MouseHouse, which is a behavioral and sensor suite for mouse experiments. Together, these tools provide a comprehensive behavioral platform for acquiring and analyzing video, sensor, and behavioral data. The work is open source and provided as a resource for the field.
Strengths:
(1) Squeakpose Studio was relatively easy to install and use. We were impressed that we were able to install it and test our own videos with minimal struggles. The authors provide installation tutorial videos that were very helpful.
(2) The GUI environment for SqueakPose Studio was very usable, and the authors should be commended on the time and effort that went into improving the useability of their system. The keypoint and skeleton configuration was flexible, allowing us to define custom body part sets without modifying code directly. The pose estimation accuracy on our own videos was good right out of the box, without requiring fine-tuning or retraining. For a tool being evaluated for the first time, this was all very impressive!
We thank this reviewer for the positive comments and have provided responses to the specific potential weaknesses noted below.
Weaknesses:
(1) While we were able to install and test Squeakpose Studio, it was not entirely seamless. The primary installation resource is a tutorial video, and we would recommend supplementing this with a written installation checklist that explicitly lists all required software dependencies (e.g. Python, UV, Visual Studio). The tutorial video was also at times unclear in distinguishing required from optional components. For example, Visual Studio is described as not necessary, yet the tutorial demonstrates the workflow entirely within that environment, so it may be challenging for a user to follow along without that. We recommend that the authors adopt a stricter, step-by-step installation guide that is prescriptive about required software and leaves little room for confusion.
We thank the reviewer for this helpful feedback and agree that the installation workflow should distinguish more clearly between required and optional components. Our goal with SqueakPose Studio is to place as much functionality as possible in the GUI so that users are not required to rely on command-line tools for additional features or advanced use. For that reason, the command-line surface is intentionally minimal: after the repository is cloned and the UV-managed environment is created, almost all functionality is accessed through the graphical interface.
We also appreciate the opportunity to clarify the point about Visual Studio. The tutorial video demonstrates the workflow using Visual Studio Code, not Visual Studio. Visual Studio Code is optional and is used in the video only as a convenient editor and interface for demonstrating the workflow. The GUI can also be launched directly from a terminal, and users may use any preferred editor or IDE, including VS Code, Zed, Cursor, Jupyter-based workflows, or no IDE at all.
We have updated the written README and YouTube walkthrough to make this distinction clearer. Specifically, provided a stricter installation checklist that separates required components, such as Python and UV, from optional tools, such as VS Code or other editors. We also demonstrated launching SqueakPose Studio directly from a terminal so users can follow the workflow without relying on a specific IDE.
(2) The paper also describes SqueakView and MouseHouse. Unfortunately, we were unable to evaluate these components as both require the MouseHouse hardware platform. Even without directly using MouseHouse, we noticed some incompleteness here, as we could not locate a bill of materials, component pricing, or assembly guide in the paper or associated GitHub repositories. Given that affordability and accessibility are central claims, a consolidated parts list, approximate costs, and a build guide or video would be necessary for most labs to realistically decide whether they plan to replicate the hardware and evaluate this functionality that the paper describes. In this regard, we felt that MouseHouse and potentially SqueakView were not sufficiently documented for publication.
We agree with the reviewer that MouseHouse and SqueakView are more difficult to evaluate than SqueakPose Studio because they involve dedicated hardware, including an edge-compute platform. This is an unavoidable tradeoff for a system designed not only for offline pose estimation, but also for real-time acquisition and deployment. We recognize, however, that if the manuscript emphasizes affordability and accessibility, then users need a clear way to estimate cost, order components, assemble the system, and reproduce the hardware configuration.
We have therefore added a consolidated bill of materials to the GitHub repository, including component names, approximate pricing, and suggested sources where appropriate. We now provide a complete guide for connecting the hardware and flashing the required firmware/software to the devices. This documentation makes clearer what is required for MouseHouse-specific functionality versus what can be used independently through SqueakPose Studio.
We also note that edge-compute devices such as the Jetson Orin Nano are increasingly common in robotics and real-time computer-vision applications, but we appreciate that many behavioral neuroscience laboratories may not yet have this hardware in place. For some users, this paper may be their first exposure to this compute platform. For that reason, we agree that the repository should provide more complete onboarding materials for labs that wish to adopt the hardware ecosystem, and we now provide that.
(3) The benchmarking comparison to DeepLabCut (DLC) introduced multiple challenges that left us unclear if the head-to-head comparison was appropriate as described. First, the dataset used for benchmarking was small and homogeneous, from the methods they used "10 min open-field tasks of single mice with bilateral photometry cables." As such, the claims about comparisons between SqueakPose Studio and DLC may be too broad, given this single test case. Specifically, this dataset does not test robustness across lighting conditions, coat colors, species, occlusions, different-shaped arenas, etc. Second, the comparison to DLC in Figure 1 does not include any quantitative statistical comparisons, which are needed to evaluate the claims that were made. For instance, the error in Figure 1e looks worse for their system than DLC, although statistical comparisons were not made. Third, there are many settings and optimizations that can be made for both systems. Without more detail, this makes it hard to know if the head-to-head comparison is really fair. Fourth - the metrics are given as very specific numbers from single runs, i.e., an inference time of 71.59 minutes in Figure 1d. This metric would be more meaningful if it reported the mean of multiple runs, with error estimation. Finally, while the code is available, the trained datasets are made available only on "reasonable request". Given the importance of these datasets to evaluating the method and allowing others to benchmark it against other systems, these should be made available on GitHub. Overall, I would recommend toning down the comparison to DLC and focusing on the strengths of Squeakpose Studio on its own merits.
We appreciate the reviewer’s thoughtful comments about the benchmarking comparison. We agree that no single dataset can establish universal performance across all lighting conditions, coat colors, species, occlusion regimes, arena geometries, or camera configurations. Our intention was not to claim that SqueakPose Studio is superior to DeepLabCut under every possible condition, nor to present a comprehensive benchmark across the full space of pose-estimation use cases. Rather, the benchmark was included as an applied demonstration of performance in a representative behavioral neuroscience workflow involving mouse open-field videos with photometry cables.
We also agree that users can substantially affect performance in any pose-estimation framework through model selection, training settings, hardware configuration, inference parameters, and optimization choices. For this reason, we view the comparison as a practical workflow benchmark rather than a definitive ranking of all possible DLC and SqueakPose Studio configurations. The primary contribution of SqueakPose Studio is not simply that it is faster in one head-to-head comparison, but that it provides an integrated GUI-based workflow for pose estimation, review, export, and real-time/edge-AI deployment.
That said, the speed improvements are not incidental. They reflect deliberate architectural and deployment choices, including the use of modern object-detection/pose-estimation architectures and optimized inference workflows. In practice, these choices can substantially reduce inference time relative to workflows that were not designed around the same deployment constraints. We will be careful in our public response and documentation not to overstate this as a universal claim across every dataset or every possible DLC configuration.
Regarding statistical comparisons and repeated runs, we agree that reporting means and variance across repeated benchmark runs can be useful. However, because this manuscript is primarily an applications and methods resource rather than a large-scale benchmarking study, we do not intend to benchmark every relevant dataset class or hardware configuration. We instead encourage users to evaluate SqueakPose Studio on their own videos and hardware, which is ultimately the most informative test for adoption in a given laboratory.
Regarding the trained datasets and models, we agree with the reviewer that broad access improves reproducibility and benchmarking. The limitation is practical rather than philosophical: the full benchmark datasets are large and are not well suited for direct hosting in a GitHub repository. We currently make these data available upon reasonable request and have included a Zenodo repo explore more appropriate public hosting options for large files, such as an institutional repository, Zenodo, OSF, or another archival data platform. We will also clarify the availability of trained models and example data so users can more easily reproduce or extend the benchmarking workflow.
Overall, we agree that SqueakPose Studio is strongest when evaluated on its own merits: accessibility, speed, GUI-based usability, flexible keypoint configuration, real-time deployment, and integration with acquisition and edge-compute workflows. We now frame the DLC comparison as a representative applied benchmark rather than as an exhaustive claim of general superiority.
(4) The paper at times makes general statements that are beyond what is shown. For instance, discussions of use in human applications are aspirational and should be treated much more conservatively in the discussion, or possibly even removed. As it stands, the discussion implies that this system can already do "zero-shot tracking of human posture and movement", enabling "a bridge between preclinical and clinical behavioral analysis". In principle, this may be true, but even for a Discussion section, this goes far beyond the capabilities that the paper actually shows.
We appreciate this comment and agree that the manuscript should distinguish more clearly between capabilities demonstrated in the present study and broader potential applications of the software architecture.
SqueakPose Studio and SqueakView are not intrinsically mouse-specific. Users can define custom classes, keypoints, and skeletons, train compatible pose-estimation models for other organisms or experimental preparations, and deploy those models using the same acquisition and inference workflow.
To make this technical capability concrete, the SqueakView repository now includes deployment-ready FP16 model packages for both the validated MouseHouse-specific pose model and a stock human-pose model. The included human-pose model demonstrates that the deployment architecture can support zero-shot human posture tracking without requiring changes to the underlying SqueakView pipeline.
We agree, however, that this technical compatibility should not be interpreted as validation for clinical behavioral analysis. The experimental demonstrations in the present manuscript focus primarily on mouse behavioral datasets. Any clinical application would require separate benchmarking, validation, and domain-specific evaluation beyond the scope of the present manuscript.
(5) While the comprehensive nature of the system and its 3 parts is impressive, I felt that it also detracted from the main focus of the paper, which was Squeakpose Studio. I might recommend dropping the other two parts, as they also require a much higher bar for a user to evaluate, and only present the Squeakpose Studio in this paper, presenting this as a general resource for pose estimation. This would also allow them more space to more comprehensively benchmark SqueakPose Studio.
We appreciate this perspective and agree that SqueakPose Studio is the most immediately accessible component of the platform for many users. However, we respectfully disagree that MouseHouse and SqueakView should be removed from the paper. The motivation for developing SqueakPose Studio was not simply to create another offline pose-estimation and analysis tool, but to enable real-time behavioral detection and deployment on edge hardware. SqueakView and MouseHouse provide the acquisition and deployment context that motivated the software architecture and demonstrate how the platform can be used in closed-loop or real-time behavioral workflows.
In developing the system, we recognized that SqueakPose Studio also functions as a user-friendly general pose-estimation interface, with features that may be useful even for laboratories that do not adopt the full MouseHouse/SqueakView ecosystem. For that reason, we presented it as both a standalone tool and as part of a broader acquisition and deployment platform.
We agree that this makes the manuscript broader than a paper focused exclusively on pose-estimation benchmarking. However, we view that breadth as important: the paper is intended to serve as a central, peer-reviewed entry point for laboratories interested in deploying real-time pose estimation in behavioral experiments. The manuscript points users to the relevant repositories, documents the design rationale, and provides a source of peer-reviewed validation for the integrated workflow. We have clarified in our response and documentation that users can adopt SqueakPose Studio independently, while MouseHouse and SqueakView support the broader real-time hardware ecosystem.
-
