Motion-Aware Heatmap Regression for Human Pose Estimation in Videos

Authors: Inpyo Song, Jongmin Lee, Moonwook Ryu, Jangwon Lee

IJCAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our motion-aware heatmap regression on Pose Track(2018, 21) and Sub-JHMDB datasets. Our results validate that the proposed motion-aware heatmaps significantly improve the precision of human pose estimation in videos, particularly in challenging scenarios such as videos like sports game footage with substantial human motions.
Researcher Affiliation Academia 1Department of Immersive Media Engineering, Sungkyunkwan University, Republic of Korea 2Electronics and Telecommunications Research Institute, Republic of Korea
Pseudocode No The paper describes its method in detail using text and mathematical formulations but does not include any pseudocode or algorithm blocks.
Open Source Code Yes (Code and related materials are available at https://github.com/ Songinpyo/MTPose.)
Open Datasets Yes Pose Track For our evaluation, we utilized two versions of the Pose Track dataset: Pose Track2018 [Iqbal et al., 2017] and Pose Track21 [Doering et al., 2022]. These datasets are key benchmarks in multi-person pose estimation and tracking within video contexts. Sub-JHMDB The Sub-JHMDB dataset [Jhuang et al., 2013] is a subset of the larger JHMDB collection.
Dataset Splits Yes We evaluated our MTPose against leading video-based human pose estimation methods on Pose Track validation sets (AP metric) and Sub-JHMDB dataset (PCK metric).
Hardware Specification Yes The training setup involves a loss weigth γ of 3/4, a batch size of 64, a learning rate of 3e-4 using the Adam W optimizer, and is conducted on a single NVIDIA Tesla V100 GPU.
Software Dependencies No The paper mentions using a Vi T Backbone and an off-the-shelf optical flow model, but does not provide specific version numbers for software dependencies like programming languages, libraries (e.g., PyTorch, TensorFlow), or other frameworks.
Experiment Setup Yes The model operates on images resized to 256 x 192 image size. We used the Vi T Backbone initialized from pretrained by [Xu et al., 2022b] and finetuned on Pose Track2018, Pose Track21 and Sub-JHMDB datasets. For the frame interval, we set it to 2 for Pose Track2018 and Sub-JHMDB, while Pose Track21 uses an interval of 1... We have set the motion threshold δ and the default standard deviation σ0 for Gaussian heatmaps at 3 for all datasets. The training setup involves a loss weigth γ of 3/4, a batch size of 64, a learning rate of 3e-4 using the Adam W optimizer...