Source-linked AI summary

SimCC: a Simple Coordinate Classification Perspective for Human Pose Estimation

Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, Shu-Tao Xia

arXiv:2107.03332v3cs.CV

TL;DR

Heatmap-based HPE suffers from quantization error, costly upsampling, refinement post-processing, and weak low-resolution performance. SimCC reformulates localization as independent horizontal and vertical classification with multiple bins per pixel, and experiments report advantages over heatmap-based counterparts while reducing computation. Its demonstrated setting is top-down pose estimation; bottom-up multi-person matching remains future work.

  • Problem

    2D heatmap HPE suffers from quantization error, costly upsampling layers, refinement post-processing, and limited low-resolution performance.

  • Method

    SimCC reformulates HPE as independent horizontal and vertical coordinate classification, uniformly dividing each pixel into multiple bins and using lightweight classifier heads.

  • Results

    SimCC shows advantages over heatmap-based representations across experiments, including higher performance with reduced computational cost in reported settings.

  • Takeaways & Limitations

    SimCC provides a simpler HPE pipeline that can omit refinement post-processing and, under certain settings, upsampling layers while supporting sub-pixel localization.

  • Takeaways & Limitations

    The paper evaluates SimCC under top-down pose estimation; bottom-up multi-person estimation faces identification ambiguity and may require extra embeddings for coordinate matching.

Abstract

from arXiv · show

The 2D heatmap-based approaches have dominated Human Pose Estimation (HPE) for years due to high performance. However, the long-standing quantization error problem in the 2D heatmap-based methods leads to several well-known drawbacks: 1) The performance for the low-resolution inputs is limited; 2) To improve the feature map resolution for higher localization precision, multiple costly upsampling layers are required; 3) Extra post-processing is adopted to reduce the quantization error. To address these issues, we aim to explore a brand new scheme, called \textit{SimCC}, which reformulates HPE as two classification tasks for horizontal and vertical coordinates. The proposed SimCC uniformly divides each pixel into several bins, thus achieving \emph{sub-pixel} localization precision and low quantization error. Benefiting from that, SimCC can omit additional refinement post-processing and exclude upsampling layers under certain settings, resulting in a more simple and effective pipeline for HPE. Extensive experiments conducted over COCO, CrowdPose, and MPII datasets show that SimCC outperforms heatmap-based counterparts, especially in low-resolution settings by a large margin.

1 Introduction

SimCC addresses quantization errors in 2D heatmap HPE by reformulating localization as horizontal and vertical coordinate classification. This design simplifies the pipeline and improves performance, particularly for low-resolution inputs.

  • 1 Introduction: 2D heatmap methods discretize continuous coordinates into downscaled maps, causing quantization error and limiting low-resolution performance.This motivates costly upsampling and refinement procedures.
  • 1 Introduction: SimCC can omit costly upsampling layers under certain settings and is presented as a simpler alternative to 1D and 2D heatmap pipelines.The comparison includes additional decoupling layers in 1D heatmap regression and refinement stages in 2D heatmaps.
  • 1 Introduction: SimCC reformulates HPE as two independent classification tasks for horizontal and vertical coordinates.A CNN- or Transformer-based backbone extracts keypoint representations before coordinate classification.
  • 1 Introduction: Uniformly dividing each pixel into multiple bins provides sub-pixel localization precision while avoiding the heatmap pipeline’s extra refinement post-processing.The method uses two lightweight classifier heads, one for each coordinate axis.
  • 1 Introduction: Over 55% GFLOPs are reduced for SimBa-Res50 while achieving higher model performance than a heatmap-based counterpart.Experiments cover COCO, CrowdPose, and MPII with different backbones and input sizes.

2 Related Work

Related work spans direct coordinate regression, heatmap-based HPE, quantization-error refinement, facial-landmark 1D heatmaps, and coordinate-classification approaches. SimCC differs by targeting HPE with a lightweight coordinate-classification pipeline.

  • 2 Related Work: Regression-based HPE directly predicts keypoint coordinates in a computationally friendly framework but has seen limited adoption because of unsatisfactory performance.
  • 2 Related Work: Heatmap-based HPE improved localization through architectures such as hourglass networks, offset aggregation, deconvolutional heads, and high-resolution representations.
  • 2 Related Work: Post-processing methods such as DARK reduce quantization error by modeling heatmap activation distributions with Taylor-expansion-based approximation.
  • 2 Related Work: Facial-landmark work decouples 2D heatmaps into 1D heatmaps using deconvolution, learnable decoupling layers, and co-attention.
  • 2 Related Work: Pix2Seq also uses classification for discrete object representations, whereas SimCC develops coordinate classification specifically for HPE and is designed to combine with CNN-based backbones.

3 SimCC: Reformulating HPE from Classification Perspective

SimCC uses a backbone and two classifier heads to predict horizontal and vertical coordinates independently. Pixel subdivision reduces quantization error and supports sub-pixel precision, while experiments examine robustness, label smoothing, and computational savings.

  • 3 SimCC: Reformulating HPE from Classification Perspective: SimCC consists of a backbone network followed by horizontal and vertical classifier heads for coordinate prediction.The backbone may be CNN- or Transformer-based, and CNN features can be flattened before classification.
  • 3 SimCC: Reformulating HPE from Classification Perspective: The splitting factor k creates Nx = W · k and Ny = H · k coordinate bins, with larger k reducing quantization error and enabling sub-pixel precision.The coordinate labels are uniformly discretized for the horizontal and vertical axes.
  • 3 SimCC: Reformulating HPE from Classification Perspective: SimCC trains coordinate classifiers with Kullback–Leibler divergence and explores equal, Laplace, and Gaussian label smoothing.Equal label smoothing is the default SimCC variant unless otherwise noted.
  • 3 SimCC: Reformulating HPE from Classification Perspective: SimCC-based methods outperform heatmap counterparts across 64×64, 128×128, 256×192, and 384×288 inputs, with especially clear gains at low resolutions.The text attributes this robustness to lower quantization error.

4 Experiments

Experiments across COCO, CrowdPose, and MPII evaluate SimCC against heatmap-based methods, with COCO results showing gains across resolutions and reduced computation. Ablations examine post-processing, upsampling, splitting factor, and label smoothing.

  • Experimental Setup: Experiments cover COCO, CrowdPose, and MPII using standard benchmark evaluations and selected CNN- and Transformer-based baselines.COCO uses average precision, while MPII uses PCKh evaluation.
  • Post-processing: SimCC omits costly refinement post-processing while maintaining higher COCO validation performance than heatmap-based counterparts across input resolutions.Heatmap methods use DARK or empirical shift, whereas SimCC operates without extra refinement post-processing.
  • COCO Results: +11.2 AP is achieved by SimCC over heatmap-based counterparts at 64×64 input resolution on COCO validation data.At 256×192, SimCC-based HRNet-W48 also improves by +0.8 AP.
  • COCO Results: +0.5 and +1.2 AP are obtained by SimCC-based HRNet-W48 and SimpleBaseline-Res50, respectively, on COCO test-dev at 384×288.The comparisons use heatmap-based counterparts.
  • Upsampling Ablation: +0.4 AP and 23.5% faster speed are achieved by deconvolution-free SimCC SimpleBaseline-Res50 versus its heatmap counterpart at 256×192.The reported speeds are 21 versus 17 FPS.
  • Ablation Study: The splitting factor k creates a trade-off: larger k lowers quantization error but makes training more difficult, so performance first increases and then decreases.Recommended settings include k = 2 for HRNet-W32 at 128×128 and 256×192, and k = 3 or 2 for SimBa-Res50 at those resolutions.
  • Ablation Study: 57.3% fewer GFLOPs are used when SimCC SimpleBaseline-Res50 removes upsampling modules while still outperforming the heatmap counterpart at 256×192.Across input resolutions, removing deconvolution layers consistently reduces computational cost.

5 Limitation and Future Work

SimCC is presented for top-down human pose estimation, while bottom-up multi-person estimation remains unresolved because multiple people create identification ambiguity.

  • SimCC introduced in this paper works under the setting of top-down human pose estimation.
  • Bottom-up multi-person pose estimation introduces identification ambiguity when multiple people are present.
  • Future work may introduce extra embeddings to address matching between candidate coordinate x and y values.

6 Conclusion

SimCC represents keypoint localization as independent horizontal and vertical classification tasks using two one-dimensional coordinate vectors. Experiments report advantages over heatmap representations and suggest implications for lightweight HPE models.

  • SimCC represents keypoint localization as two independent classification tasks for horizontal and vertical axes.
  • Joint locations are represented by two independent 1D vectors for the x- and y-coordinates.
  • Experimental results show that the 2D structure might not be essential for coordinate representations to sustain superior performance.
  • SimCC shows performance advantages over heatmap-based representations and may inspire lightweight model designs for HPE.
Loading 2107.03332v3…