Deep Human Dynamics Prior

Cui, Qiongjie; Sun, Huaijiang; Kong, Yong; Sun, Xiaoning

doi:10.1145/3474085.3475581

Cited by 4 publications

(5 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…, [95], [96], [97], [98] The fundamental distinction among the different moving average filters is the weighting function. Whereas in SMA filters the information extrapolated from each frame of the window is weighted equally, weighted moving average (WMA) filters assign different weights to different frames.…”

Section: A Moving Averagementioning

confidence: 99%

See 1 more Smart Citation

Denoising and Completion Filters for Human Motion Software: a Survey with Code

Martini,

Calanca,

Bombieri

2024

Preprint

View full text Add to dashboard Cite

Software platforms used for human motion analysis are increasingly being utilized in various fields, ranging from Healthcare to Industry 5.0. However, the inherent inaccuracy of these platforms often results in noisy descriptions of human poses or time periods during which the information is even missing. As a consequence, data filtering for denoising or completion is a common yet fundamental step before data analysis. Over the years, different techniques have been proposed, from generalpurpose solutions based on low-pass filters to more advanced and embedded approaches based on state observers rather than deep learning. This survey presents the current state-of-the-art filtering solutions for denoising and completing data generated by software platforms for human motion analysis, with a focus on 3D positional data extrapolated through marker-based or markerless motion capture systems. It proposes a concise taxonomy based on the filter technology and application assumptions. For each class, it summarizes the basic concepts and reports application feedback collected from literature. The survey also includes implementation codes or links to the authors’ original codes, allowing readers to easily reproduce all the algorithms in different experimental settings (https://github.com/PARCO-LAB/mocap-refinement).

show abstract

Section: A Moving Averagementioning

confidence: 99%

“…This is done by implicitly learning the structure of sk. Literature examples of DAE implementations are [79], [80], [78], [81], [82], [83], [84], [85], [95], [86], [88], [87], [89], [90], [91], [92].…”

Section: B Denoising Autoencodermentioning

confidence: 99%

Denoising and Completion Filters for Human Motion Software: a Survey with Code

Martini,

Calanca,

Bombieri

2024

Preprint

View full text Add to dashboard Cite

show abstract

“…While marker-free, image-based solutions have yielded promising results, markerbased motion capture remains popular due to its accuracy and flexibility [Taheri et al 2020]. MoCap systems' high-fidelity body motion data [CMU 2000;Mahmood et al 2019;Sigal et al 2010;Trumble et al 2017] is useful in a variety of applications [Zheng et al 2020], including action recognition [Hua et al 2023;Yan et al 2018], action prediction [Cao et al 2020;Cui et al 2021], motion synthesis Tevet et al 2022] and image-based pose estimation [Munea et al 2020;Varol et al 2017]. However, raw MoCap data is inevitably contaminated with errors.…”

Section: Related Workmentioning

confidence: 99%

A Locality-based Neural Solver for Optical Motion Capture

Pan,

Zheng,

Jiang

et al. 2023

SIGGRAPH Asia 2023 Conference Papers

View full text Add to dashboard Cite

We present a novel locality-based learning method for cleaning and solving optical motion capture data. Given noisy marker data, we propose a new heterogeneous graph neural network which treats markers and joints as different types of nodes, and uses graph convolution operations to extract the local features of markers and joints and transform them to clean motions. To deal with anomaly markers (e.g. occluded or with big tracking errors), the key insight is that a marker's motion shows strong correlations with the motions of its immediate neighboring markers but less so with other markers, a.k.a. locality, which enables us to efficiently fill

show abstract

“…Human Motion Forecasting. Nowadays, various GNNbased models are being developed to extract the semantic connectivity of the 3D skeleton sequence, with promising results (Mao et al 2019;Cui et al 2020;Li et al 2020a;Dang et al 2021;Cui et al 2021b;Zhong et al 2022). However, GCNs are capable only of gathering information from the local neighbor joints, and have a limited capacity to capture long-term relationships.…”

Section: Related Workmentioning

confidence: 99%

“…Researchers typically train on external large-scale datasets (Ionescu et al 2014) to achieve a generic pre-trained model, which is then indiscriminately applied to all test sequences with the same set of network weights in the inference stage (Jain et al 2016;Wei Mao 2021;Dang et al 2021). These approaches have extensively investigated this issue from various perspectives, emerging as the mainstream solutions (Guo et al 2022;Cui et al 2021b).…”

Section: Introductionmentioning

confidence: 99%

Meta-Auxiliary Learning for Adaptive Human Pose Prediction

Cui

Sun

et al. 2023

AAAI

View full text Add to dashboard Cite

Predicting high-fidelity future human poses, from a historically observed sequence, is crucial for intelligent robots to interact with humans. Deep end-to-end learning approaches, which typically train a generic pre-trained model on external datasets and then directly apply it to all test samples, emerge as the dominant solution to solve this issue. Despite encouraging progress, they remain non-optimal, as the unique properties (e.g., motion style, rhythm) of a specific sequence cannot be adapted. More generally, once encountering out-of-distributions, the predicted poses tend to be unreliable. Motivated by this observation, we propose a novel test-time adaptation framework that leverages two self-supervised auxiliary tasks to help the primary forecasting network adapt to the test sequence. In the testing phase, our model can adjust the model parameters by several gradient updates to improve the generation quality. However, due to catastrophic forgetting, both auxiliary tasks typically have a low ability to automatically present the desired positive incentives for the final prediction performance. For this reason, we also propose a meta-auxiliary learning scheme for better adaptation. Extensive experiments show that the proposed approach achieves higher accuracy and more realistic visualization.

show abstract

Deep Human Dynamics Prior

Cited by 4 publications

References 30 publications

Denoising and Completion Filters for Human Motion Software: a Survey with Code

Denoising and Completion Filters for Human Motion Software: a Survey with Code

A Locality-based Neural Solver for Optical Motion Capture

Meta-Auxiliary Learning for Adaptive Human Pose Prediction

Contact Info

Product

Resources

About