Pith. sign in

REVIEW 3 major objections 6 minor 42 references

ProGait: A Multi-Purpose Video Dataset and Benchmark for Transfemoral Prosthesis Users

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ProGait is a 412-clip video dataset of above-knee prosthesis users, and the paper's claim is that fine-tuning standard vision models on it makes those models detect prosthetic legs and classify gait deviations far better than…

desk verdict Valuable dataset, but the benchmark claims overstate what the evidence shows, and the gait classification protocol is arithmetically impossible as described. read the letter →

arxiv 2507.10223 v1 pith:7GU7JCOC submitted 2025-07-14 cs.CV cs.AI

Xiangyu Yin , Boyuan Yang , Weichen Liu , Qiyao Xue , Abrar Alamri , Goeran Fiedler , Wei Gao This is my paper
classification cs.CVcs.AI
keywords ProGaitdatasettransfemoralprosthesisanalysisvideoobjectsegmentation2Dhumanposeestimationclassificationprostheticbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ProGait is a video dataset built from 412 clips of four above-knee (transfemoral) prosthesis users walking on newly fitted prosthetic legs, recorded from frontal and sagittal views in two clinical scenarios. The paper claims that this dataset lets vision models do three things generic models do poorly on prosthetic limbs: segment the walking person, estimate 2D body pose including the prosthetic knee and foot, and classify nine clinical gait-deviation categories. The authors report that fine-tuning standard models on ProGait improves segmentation and pose estimation substantially, with RTMPose-ProGait reaching AP 0.947 versus 0.855 for the pre-trained checkpoint, and that a pose-sequence LSTM classifies sagittal-view gaits at 82.6% top-1 accuracy. If the claim holds, prosthesis alignment and rehabilitation assessment could move from motion-capture labs to ordinary video.

What carries the argument

The load-bearing object is the dataset itself, built through a human-in-the-loop annotation pipeline rather than by manual tracing alone. Initial masks come from GroundingDINO and Grounded SAM2, prompted with “a human with prosthetic leg,” with additional boxes added so the prosthesis can be tracked as a separate object; the SAM2 masks are then propagated, inspected, and manually corrected when tracking fails. Initial pose keypoints come from the pre-trained RTMW model applied to the masked video; roughly 100 frames per video are manually corrected, the model is fine-tuned, and after two such stages fewer than 25% of videos require further manual correction. The benchmark protocol keeps test subjects out of the training and validation sets, and evaluations run on every frame rather than sparsely sampled frames. This pipeline is what lets the authors attribute their reported improvements to the dataset itself rather than to test-set leakage.

What would settle it

Take a held-out subset of ProGait videos and have clinicians or a motion-capture system independently label prosthetic knee, ankle, and foot keypoints and body masks; if the fine-tuned RTMPose's advantage over the pre-trained checkpoint largely disappears when scored against those independent labels rather than the RTMW-derived labels, the central claim of improved prosthetic-limb detection is not established.

Watch

Extended reading notes

Core claim

The paper's central claim is that the ProGait dataset closes a specific gap: off-the-shelf vision models, trained mostly on able-bodied people, fail to recognize prosthetic legs as part of the human body, and this failure propagates into pose estimation and gait analysis. To show the gap is closable, the authors built 412 clips of four above-knee amputees walking inside parallel bars and in open hallways, with frontal and sagittal cameras, and annotated each clip with segmentation masks, 23 whole-body keypoints (17 body, 6 foot) following the COCO-WholeBody format, and clinician-written gait descriptions organized into 9 categories. They then fine-tuned YOLO11 for segmentation and RTMPose for pose, reporting that RTMPose-ProGait reaches AP 0.947 overall and 0.918 on knee/ankle/foot keypoints, versus 0.855 and 0.804 for the stock RTMPose checkpoint, and that a pose-sequence LSTM classifies the 9 gait categories at 82.6% top-1 accuracy from sagittal views. The paper also states that its baselines outperform zero-shot state-of-the-art models by 9% in the segmentation task and 10–30% in the pose task, and it shows that several pre-trained gait-recognition models improve after fine-tuning on ProGait.

Load-bearing premise

The ground-truth labels for masks and poses were produced by pre-trained models and then manually corrected, not by an independent gold-standard measurement; if those models carry a systematic bias about how prosthetic knees and feet look, the reported benchmark improvements partly measure agreement with that bias rather than true accuracy.

Editorial extensions

If this is right

  • Fine-tuning on ProGait raises segmentation and pose performance specifically on prosthetic limbs: RTMPose-ProGait scores AP 0.947 overall and 0.918 on knee/ankle/foot keypoints, versus 0.855 and 0.804 for the stock checkpoint.
  • The subject-disjoint split means the improved test-set numbers are not from memorizing the four walkers; a held-out subject's gait is unseen by the fine-tuned model.
  • Sagittal-view pose sequences alone support gait classification (82.6% top-1 accuracy, 79.0% balanced accuracy), while mixing frontal and sagittal sequences hurts accuracy, pointing to sagittal video as the informative viewpoint.
  • Lower-body keypoints alone retain most classification performance (77.3% sagittal top-1 accuracy), supporting a reduced-input setup for downstream gait analysis.
  • Established gait-recognition models improve when fine-tuned on ProGait (for example, GaitGraph2 rises from 0.200 to 0.440 top-1 accuracy), showing the dataset transfers beyond the paper's own baselines.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported gains are measured against labels that RTMW and SAM2 helped create, so part of the fine-tuned models' advantage may be learning the label generator's biases; testing on independently obtained motion-capture or expert labels would reveal how much is genuine.
  • Because the dataset holds only four subjects, all in the same age band and amputation etiology, the benchmark numbers likely overstate how well the fine-tuned models would transfer to younger or trauma-related amputees; adding subjects outside this profile is the natural next test.
  • The pairing of clinician-written gait descriptions with video could support an LLM-based explainable alignment assistant that names the deviation and suggests a prosthesis adjustment, a use the paper flags as future work but does not evaluate.
  • The strong sagittal-view result suggests a single side camera, rather than two views, may be enough for automated gait classification, which would simplify clinical deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces ProGait, a new video dataset containing 412 clips of four above-knee amputees walking with multiple prosthetic legs in two scenarios (inside and outside parallel bars), along with annotations for video object segmentation (masks), 2D human pose estimation (23 body and foot keypoints), and gait analysis (9 coarse gait categories with expert textual descriptions). The authors provide fine-tuned baselines (YOLO11 for VOS, RTMPose for HPE, and a custom LSTM for gait classification) and compare them against pre-trained or zero-shot SOTA models. The central claim is that using ProGait improves the generalizability of vision models for prosthesis-specific tasks, supported by quantitative comparisons in Tables 3-8. The dataset and code are publicly available.

Significance. ProGait addresses a real gap: existing vision datasets largely exclude individuals with prosthetic limbs, and specialized prosthetic gait datasets often lack raw video and multi-task annotations. The dataset's design—paired frontal/sagittal views, segmentation masks, keypoints, and expert textual gait assessments—is well suited to support clinically relevant vision tasks. The public release of code and dataset is a strength, and the paper provides quantitative benchmarks across three tasks. However, the quantitative support for the headline generalization claim is currently weakened by an internal inconsistency in the gait classification evaluation protocol and by a contradiction between the abstract's VOS claim and Table 3. If these are fixed, the dataset and benchmarks could be a useful community resource.

major comments (3)
  1. [§5.4, Table 6 (cf. §4.2 and §5.1)] The paper states in §4.2 and §5.1 that test subjects are excluded from training and validation, yet Table 6 reports a 5-fold cross-validation with only four subjects (P1–P4). A 5-fold subject-disjoint split is arithmetically impossible with four subjects; at most four non-empty held-out subject groups can be formed. The reported cross-validation is therefore either clip-level (same subject appears in training and test) or the protocol statement is false. With clip-level splits, the LSTM can exploit person-specific gait signatures, so the sagittal top-1 accuracy of 0.826 may largely reflect subject identification rather than classification of the nine gait categories. This directly undermines the paper's headline claim of improved generalizability. Please re-run the evaluation with leave-one-subject-out (4 folds), report per-subject results, or explicitly state that the split is clip-level and temper the generalizability claim accordingly.
  2. [§1 and Abstract vs. Table 3] The introduction claims that baseline models 'outperform the SOTA models by 9% in VOS task.' Table 3 shows that Grounded SAM2 with the prompt 'a human with prosthetic leg.' achieves mIoU 0.964, whereas the fine-tuned YOLO11-ProGait achieves 0.847. Against that SOTA model, the baseline is worse by about 12 mIoU points, not better by 9%. If the intended comparison is against the original YOLO11 checkpoint (0.784), the 9% figure is roughly consistent, but the text and Figure 1 present Grounded SAM2 as the SOTA reference. Please clarify which comparison is meant and correct the abstract and introduction to match Table 3.
  3. [§3.2 and §5.3] The pose ground truth is produced by a fine-tuned RTMW model, with manual correction on fewer than 25% of videos after two fine-tuning stages. The authors appropriately refrain from using RTMW as the HPE baseline (§4.2), but the test-set labels are still partly model-derived. If systematic errors remain on prosthetic knee and foot keypoints, the AP improvements in Tables 4 and 5 (e.g., 0.947 vs. 0.855) could overstate true pose quality. Please report the fraction of test frames that received manual correction, provide per-keypoint error analysis on the prosthetic limb, and ideally validate a subset of ground-truth poses against independent annotations or motion capture.
minor comments (6)
  1. [Abstract and §1 vs. Table 6] The abstract and introduction report gait classification accuracy as 'up to 81.2%' for sagittal views, but Table 6 reports 0.826 (82.6%) for the sagittal view; the numbers should be harmonized.
  2. [Table 2] The class distribution in Table 2 sums to 144, which equals the number of walking trials, but the dataset contains 412 video clips; clarify whether gait labels are assigned per trial or per clip, and report clip-level class counts for the gait classification benchmark.
  3. [§5.4] The phrase 'we conducted a 5-fold cross-validation and evaluated the model across five different setups' is ambiguous; it should state explicitly whether the five rows of Table 6 are the folds or the evaluation configurations.
  4. [§4.1, Eq. (3)] In the OKS definition, the indicator function δ(vi > 0) should be δ(vi = 1) if visibility is encoded as a binary flag; please correct the notation to align with the COCO-WholeBody convention.
  5. [Table 5] There is a stray period after '0.814' in the RTMPose row under AP-leg (outside); this appears to be a typo.
  6. [§3.2] The sentence 'we apply it to the original videos and estimate the poses by applying the pre-trained RTMW model within the MMPose framework on the masked video' is grammatically awkward and should be rephrased for clarity.

Circularity Check

2 steps flagged · score 5.0 of 10

HPE benchmark is partially self-referential because test-set pose labels are RTMW-derived, and the 5-fold gait CV cannot be subject-disjoint with only four subjects, undermining the generalizability claim; the dataset itself remains an independent contribution.

  1. self definitional [§3.2 Annotations, §4.2 Baseline Models, §5.3 HPE results]
    "Since the majority of our pose annotations, including those in the test set, are derived from the fine-tuned RTMW model, we refrain from using it as the baseline for the 2D Human Pose Estimation task."

    The HPE ground-truth keypoints are not independent measurements; they are outputs of RTMW, a pose-estimation model, with manual correction applied only until <25% of videos needed further correction. RTMPose-ProGait is fine-tuned on this same label distribution and then evaluated against it, so the reported AP gain (0.947 vs 0.855) partly measures agreement with RTMW's labeling biases rather than anatomical accuracy. The target variable is thus defined by a model of the same task the benchmark claims to test, making the improvement partially circular.

  2. other [§5.4 Gait Classification; §5.1 experimental setup; Table 1 subject counts]
    "Since no off-the-shelf model exists for this specific task, we conducted a 5-fold cross-validation and evaluated the model across five different setups, as shown in Table 6."

    With only four subjects (P1–P4 in Table 1), a 5-fold subject-disjoint split is arithmetically impossible, yet §5.1 promises that 'all the subjects in the test set are not present in either the training set or the validation set.' Therefore the folds must share subject identity, allowing the LSTM to exploit person-specific gait signatures. The sagittal top-1 of 0.826 (and the intro's 81.2%) may then reflect subject identification rather than classification of the nine gait categories, so the headline 'improved generalizability' is not supported by this protocol. This is a leakage/validity flaw rather than a definitional reduction, but it closes the evaluation loop through subject identity.

full rationale

ProGait is primarily a dataset contribution, and much of its content is independent: the videos, expert gait-category labels, and segmentation masks (SAM2 with manual refinement) are external to any derivation. No parameter is fitted and then renamed a prediction, and no load-bearing self-citation is used. The circularity is confined to the benchmark claims. For HPE, the paper explicitly acknowledges that test-set pose annotations are RTMW-derived; evaluating a model fine-tuned on those same labels against them is partially self-referential, though manual correction and the COCO-WholeBody keypoint schema provide some independent grounding. For gait classification, the 5-fold CV with only four subjects cannot be subject-disjoint as promised, so the reported accuracy is vulnerable to subject-identity leakage; this is an evaluation-validity problem adjacent to circularity rather than a derivation-equivalence problem. The VOS comparison is also weakened by the intro's 9% claim being inconsistent with Table 3, but that is a reporting issue, not circularity. Overall, the dataset itself stands on independent evidence, but the two headline performance claims contain partial closed loops, warranting a moderate score.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper makes no parametric derivation; its central claim is empirical. The main unverified inputs are the reliability of model-derived annotations as ground truth and the clinical validity of the nine gait categories. These are domain assumptions rather than free parameters.

assumptions (3)
  • domain assumption Model-generated annotations with manual correction are accurate enough to serve as ground truth for benchmarking.
    Section 3.2: SAM2 and RTMW generate masks and keypoints; pose keypoints are auto-annotated after two fine-tuning rounds, with <25% of videos needing further manual correction. No independent validation against motion capture or clinical measurement is reported.
  • domain assumption The nine gait categories and their per-sample assignments reflect clinically meaningful, separable classes.
    Appendix A defines categories; Table 2 shows an imbalanced distribution with classes as small as 4 samples, which makes the reported accuracy partly a property of the small labeled set.
  • domain assumption Subjects with vascular-related amputation are representative of the broader transfemoral prosthesis population.
    Section 3.1 states all subjects are middle-aged or elderly above-knee amputees with vascular issues; generalization to other etiologies and age groups is unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProGait: A Multi-Purpose Video Dataset and Benchmark for Transfemoral Prosthesis Users." pith.science (2026). https://pith.science/paper/7GU7JCOC

@misc{pith2026250710223,
  author       = {Pith},
  title        = {Pith review of: ProGait: A Multi-Purpose Video Dataset and Benchmark for Transfemoral Prosthesis Users},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7GU7JCOC}},
  note         = {Machine review of arXiv:2507.10223}
}
read the original abstract

Prosthetic legs play a pivotal role in clinical rehabilitation, allowing individuals with lower-limb amputations the ability to regain mobility and improve their quality of life. Gait analysis is fundamental for optimizing prosthesis design and alignment, directly impacting the mobility and life quality of individuals with lower-limb amputations. Vision-based machine learning (ML) methods offer a scalable and non-invasive solution to gait analysis, but face challenges in correctly detecting and analyzing prosthesis, due to their unique appearances and new movement patterns. In this paper, we aim to bridge this gap by introducing a multi-purpose dataset, namely ProGait, to support multiple vision tasks including Video Object Segmentation, 2D Human Pose Estimation, and Gait Analysis (GA). ProGait provides 412 video clips from four above-knee amputees when testing multiple newly-fitted prosthetic legs through walking trials, and depicts the presence, contours, poses, and gait patterns of human subjects with transfemoral prosthetic legs. Alongside the dataset itself, we also present benchmark tasks and fine-tuned baseline models to illustrate the practical application and performance of the ProGait dataset. We compared our baseline models against pre-trained vision models, demonstrating improved generalizability when applying the ProGait dataset for prosthesis-specific tasks. Our code is available at https://github.com/pittisl/ProGait and dataset at https://huggingface.co/datasets/ericyxy98/ProGait.

Figures

Figures reproduced from arXiv: 2507.10223 by the authors.

Figure 1
Figure 1. Limitation of current vision-based ML models on detect [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Examples of video frames and annotations from the ProGait dataset. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The two scenarios ditions.1 Each walking trial includes both frontal and sagittal views, providing comprehensive perspectives for analysis. To ensure diversity and generalizability, the trials on each subject involve various types and configurations of pros￾thetic legs, different background contexts and lighting con￾ditions, and heterogeneous presence of other human indi￾viduals. The dataset covers a diverse range o… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Components of a transfemoral prosthetic leg (left) and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Video Recording Setup of knee and ankle joints, and different lengths of the pylon. Such differences affect the subject’s gait patterns in differ￾ent ways, and also result in very distinct visual appearances that make it hard for the vision models to recognize. To capt…
Figure 7
Figure 7. Figure 7: Zero-shot pose estimation with RTMW keypoints. We then fine-tune the RTMW model on these annotated frames and use the fine-tuned model to infer key￾points for a larger set of frames (∼1,000). Next, we visually inspect the annotations from the fine-tuned model and se￾le…
Figure 8
Figure 8. Figure 8: Gait annotations Gait Analysis (GA). To facilitate gait analysis, we engaged researchers in rehabilitation sciences and human engineer￾ing to provide detailed textual descriptions for each video sample. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: The pipeline of baseline models 4.2. Baseline Models To benchmark our dataset in the aforementioned tasks, we fine-tuned several vision models listed below as baselines. These models operate in a pipeline, as shown in [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Fine-tuned YOLO may detect subject body as separate [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: The sequence of pose keypoint coordinates [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 12
Figure 12. Figure 12: Videos and corresponding annotations captured at [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Videos and corresponding annotations captured at [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Videos and corresponding annotations captured at [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Videos and corresponding annotations captured at [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 37 canonical work pages

  1. [1]

    Epidemiology and risk of amputation in patients with diabetes mellitus and peripheral artery disease

    J Aaron Barnes, Mark A Eid, Mark A Creager, and Philip P Goodney. Epidemiology and risk of amputation in patients with diabetes mellitus and peripheral artery disease. Arte- riosclerosis, thrombosis, and vascular biology, 40(8):1808– 1817, 2020. 3 9

    work page 2020

  2. [2]

    Small n designs for rehabilita- tion research

    Scott D Barnett, Allen W Heinemann, Alexander Libin, Arthur C Houts, Julie Gassaway, Sunil Sen-Gupta, Aaron Resch, and Daniel F Brossart. Small n designs for rehabilita- tion research. Journal of Rehabilitation Research & Devel- opment, 49(1), 2012. 4

    work page 2012

  3. [3]

    Appropriateness of sample sizes in published research on prosthetic knee com- ponentry

    Rachel Beaudette and Goeran Fiedler. Appropriateness of sample sizes in published research on prosthetic knee com- ponentry. JPO: Journal of Prosthetics and Orthotics, 30(2): 60–68, 2018. 4

    work page 2018

  4. [4]

    Force plate analysis of the walking gait in healthy dogs

    Steven C Budsberg, Mary C Verstraete, and Robert W Soutas-Little. Force plate analysis of the walking gait in healthy dogs. American journal of veterinary research , 48 (6):915–918, 1987. 3

    work page 1987

  5. [5]

    The effect of pros- thetic alignment on the stump temperature and ground re- action forces during gait in transfemoral amputees

    Andres M C ´ardenas, Juliana Uribe, Josep M Font-Llagunes, Alher M Hern ´andez, and Jes ´us A Plata. The effect of pros- thetic alignment on the stump temperature and ground re- action forces during gait in transfemoral amputees. Gait & Posture, 95:76–83, 2022. 1

    work page 2022

  6. [6]

    Com- parison of markerless and marker-based motion capture tech- nologies through simultaneous data collection during gait: proof of concept

    Elena Ceseracciu, Zimi Sawacha, and Claudio Cobelli. Com- parison of markerless and marker-based motion capture tech- nologies through simultaneous data collection during gait: proof of concept. PloS one, 9(3):e87640, 2014. 1, 3

    work page 2014

  7. [7]

    Validation of portable in-clinic video-based gait analysis for prosthesis users

    Anthony Cimorelli, Ankit Patel, Tasos Karakostas, and R James Cotton. Validation of portable in-clinic video-based gait analysis for prosthesis users. Scientific Reports, 14(1): 3840, 2024. 1, 3

    work page 2024

  8. [8]

    Benchmarking of a full- body inertial motion capture system for clinical gait analysis

    Teunis Cloete and Cornie Scheffer. Benchmarking of a full- body inertial motion capture system for clinical gait analysis. In 2008 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society , pages 4579–

    work page 2008

Show all 42 references
  1. [9]

    Openmmlab pose estimation tool- box and benchmark

    MMPose Contributors. Openmmlab pose estimation tool- box and benchmark. https://github.com/open- mmlab/mmpose, 2020. 5

    2020

  2. [10]

    Gait analysis in lower-limb amputation and prosthetic rehabilitation

    Alberto Esquenazi. Gait analysis in lower-limb amputation and prosthetic rehabilitation. Physical Medicine and Reha- bilitation Clinics, 25(1):153–167, 2014. 2

    2014

  3. [11]

    Opengait: Revisiting gait recognition towards better practicality

    Chao Fan, Junhao Liang, Chuanfu Shen, Saihui Hou, Yongzhen Huang, and Shiqi Yu. Opengait: Revisiting gait recognition towards better practicality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9707–9716, 2023. 9

    2023

  4. [12]

    Gpgait: Generalized pose-based gait recognition

    Yang Fu, Shibei Meng, Saihui Hou, Xuecai Hu, and Yongzhen Huang. Gpgait: Generalized pose-based gait recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19595–19604, 2023. 9

    2023

  5. [13]

    The effect of risk and race on lower extrem- ity amputations among medicare diabetic patients

    Joshua B Goldberg, Philip P Goodney, Jack L Cronenwett, and Frank Baker. The effect of risk and race on lower extrem- ity amputations among medicare diabetic patients. Journal of vascular surgery, 56(6):1663–1668, 2012. 3

    2012

  6. [14]

    National trends in lower ex- tremity bypass surgery, endovascular interventions, and ma- jor amputations

    Philip P Goodney, Adam W Beck, Jan Nagle, H Gilbert Welch, and Robert M Zwolak. National trends in lower ex- tremity bypass surgery, endovascular interventions, and ma- jor amputations. Journal of vascular surgery, 50(1):54–60,

  7. [15]

    Issues affecting the level of prosthetics research evidence: Secondary analysis of a systematic review

    Brian J Hafner and Andrew B Sawers. Issues affecting the level of prosthetics research evidence: Secondary analysis of a systematic review. Prosthetics and orthotics international, 40(1):31–43, 2016. 4

    2016

  8. [16]

    A proper sequence of dynamic alignment in transtibial prosthesis: insight through socket reaction mo- ments

    Hiroshi Hashimoto, Toshiki Kobayashi, Fan Gao, and Masa- taka Kataoka. A proper sequence of dynamic alignment in transtibial prosthesis: insight through socket reaction mo- ments. Scientific reports, 13(1):458, 2023. 7

    2023

  9. [17]

    A kinematic and ki- netic dataset of 18 above-knee amputees walking at various speeds

    Sarah Hood, Marshall K Ishmael, Andrew Gunnell, KB Foreman, and Tommaso Lenzi. A kinematic and ki- netic dataset of 18 above-knee amputees walking at various speeds. Scientific data, 7(1):150, 2020. 3

    2020

  10. [18]

    Rtmpose: Real- time multi-person pose estimation based on mmpose

    Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Real- time multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 3, 7, 8

    2023 arXiv

  11. [19]

    Rtmw: Real-time multi-person 2d and 3d whole-body pose estimation

    Tao Jiang, Xinchen Xie, and Yining Li. Rtmw: Real-time multi-person 2d and 3d whole-body pose estimation. arXiv preprint arXiv:2407.08634, 2024. 5

    2024 arXiv

  12. [20]

    Whole-body human pose estimation in the wild

    Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In European Conference on Computer Vision, pages 196–214. Springer, 2020. 5, 6, 8

    2020

  13. [21]

    Ultralytics yolo11, 2024

    Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 1, 3, 7

    2024

  14. [22]

    The effect of alignment changes on uni- lateral transtibial amputee’s gait: a systematic review

    Niels Jonkergouw, Maarten R Prins, Arjan WP Buis, and Pe- ter van der Wurff. The effect of alignment changes on uni- lateral transtibial amputee’s gait: a systematic review. Plos one, 11(12):e0167466, 2016. 7

    2016

  15. [23]

    Computer vision methods for parkinsonian gait analysis: A review on patents

    Taha Khan, Peter Grenholm, and Dag Nyholm. Computer vision methods for parkinsonian gait analysis: A review on patents. Recent Patents on Biomedical Engineering (Discon- tinued), 6(2):97–108, 2013. 1, 3

    2013

  16. [24]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 3

    2023

  17. [25]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision , pages 38–55. Springe...

    2024

  18. [26]

    A mobile force plate and three-dimensional motion analysis system for three-dimensional gait assessment

    Tao Liu, Yoshio Inoue, Kyoko Shibata, and K Shiojima. A mobile force plate and three-dimensional motion analysis system for three-dimensional gait assessment. IEEE Sensors Journal, 12(5):1461–1467, 2011. 3

    2011

  19. [27]

    Comparative abilities of microsoft kinect and vicon 3d motion capture for gait analysis

    Alexandra Pfister, Alexandre M West, Shaw Bronner, and Jack Adam Noah. Comparative abilities of microsoft kinect and vicon 3d motion capture for gait analysis. Journal of medical engineering & technology, 38(5):274–280, 2014. 1, 3

    2014

  20. [28]

    De- sign optimization in lower limb prostheses: A review

    Mark A Price, Philipp Beckerle, and Frank C Sup. De- sign optimization in lower limb prostheses: A review. IEEE Transactions on Neural Systems and Rehabilitation Engi- neering, 27(8):1574–1588, 2019. 2

    2019

  21. [29]

    Computer vision for clinical gait analysis: A gait abnormality video dataset

    Rahm Ranjan, David Ahmedt-Aristizabal, Mohammad Ali Armin, and Juno Kim. Computer vision for clinical gait analysis: A gait abnormality video dataset. arXiv preprint arXiv:2407.04190, 2024. 3 10

    2024

  22. [30]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5

    2024 arXiv

  23. [31]

    You only look once: Unified, real-time object de- tection

    J Redmon. You only look once: Unified, real-time object de- tection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 1, 3

    2016

  24. [32]

    Grounded sam: Assembling open-world models for diverse visual tasks

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 ,

    arXiv

  25. [33]

    Energy expenditure and biomechanical characteristics of lower limb amputee gait:: The influence of prosthetic align- ment and different prosthetic components

    Thomas Schmalz, Siegmar Blumentritt, and Rolf Jarasch. Energy expenditure and biomechanical characteristics of lower limb amputee gait:: The influence of prosthetic align- ment and different prosthetic components. Gait & posture, 16(3):255–263, 2002. 1

    2002

  26. [34]

    Vision-based gait recognition: A sur- vey

    Jasvinder Pal Singh, Sanjeev Jain, Sakshi Arora, and Uday Pratap Singh. Vision-based gait recognition: A sur- vey. Ieee Access, 6:70497–70527, 2018. 1, 3

    2018

  27. [35]

    Deep high-resolution representation learning for human pose es- timation

    Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose es- timation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8

    2019

  28. [36]

    Towards a deeper under- standing of skeleton-based gait recognition

    Torben Teepe, Johannes Gilg, Fabian Herzog, Stefan H¨ormann, and Gerhard Rigoll. Towards a deeper under- standing of skeleton-based gait recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1569–1577, 2022. 9

    2022

  29. [37]

    Vipnas: Ef- ficient video pose estimation via neural architecture search

    Lumin Xu, Yingda Guan, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Vipnas: Ef- ficient video pose estimation via neural architecture search. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 16072–160...

    2021

  30. [38]

    Vit- pose: Simple vision transformer baselines for human pose estimation

    Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vit- pose: Simple vision transformer baselines for human pose estimation. Advances in neural information processing sys- tems, 35:38571–38584, 2022. 8

    2022

  31. [39]

    Health & gait: a dataset for gait-based analysis

    Jorge Zafra-Palma, Nuria Mar ´ın-Jim´enez, Jos ´e Castro- Pi˜nero, Magdalena Cuenca-Garc ´ıa, Rafael Mu ˜noz-Salinas, and Manuel J Mar ´ın-Jim´enez. Health & gait: a dataset for gait-based analysis. Scientific Data, 12(1):44, 2025. 3

    2025

  32. [40]

    Alignment of l owe-limb prostheses

    M Zahedi, W Spence, S Solomonidis, and J Paul. Alignment of l owe-limb prostheses. J Rehabil Res Dev , 23(2):2–19,

  33. [41]

    Effect of prosthetic alignment on gait and biomechanical loading in individuals with transfemoral amputation: A preliminary study

    Tengyu Zhang, Xuefei Bai, Fei Liu, and Yubo Fan. Effect of prosthetic alignment on gait and biomechanical loading in individuals with transfemoral amputation: A preliminary study. Gait & Posture, 71:219–226, 2019. 1

    2019

  34. [42]

    Gait patterns as biomarkers: A video- based approach for classifying scoliosis

    Zirui Zhou, Junhao Liang, Zizhao Peng, Chao Fan, Fengwei An, and Shiqi Yu. Gait patterns as biomarkers: A video- based approach for classifying scoliosis. In International Conference on Medical Image Computing and Computer- Assisted Intervention, pages 284–294. Springer, 2024....

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.