Abstract:In this study we provide empirical evidence demonstrating that the quality of training data impacts model performance in Human Pose Estimation (HPE). Inaccurate labels in widely used data sets, ranging from minor errors to severe mislabeling, can negatively influence learning and distort performance metrics. We perform an in-depth analysis of popular HPE data sets to show the extent and nature of label inaccuracies. Our findings suggest that accounting for the impact of faulty labels will facilitate the development of more robust and accurate HPE models for a variety of real-world applications. We show improved performance with cleansed data.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to explore and solve the impact of label errors in the training dataset on the performance of the **Human Pose Estimation (HPE)** model. Specifically, the paper focuses on the following aspects: 1. **Prevalence and impact of label errors**: - Research shows that there are various label errors ranging from minor to severe mis - labeling in widely - used HPE datasets, and these errors will negatively affect the model's learning effect and distort performance evaluation metrics. - The paper demonstrates the scope and nature of these label errors through empirical analysis, emphasizing the importance of identifying and dealing with these problems. 2. **Impact of data quality issues on model performance**: - Label errors in the dataset will lead to a decline in the model's training and generalization performance. The paper provides experimental evidence that using cleaned data can significantly improve model performance. - Especially in real - world applications, ensuring data quality is crucial for developing more robust and accurate HPE models. 3. **Label error classification and detection methods**: - The paper proposes a detailed label error classification (taxonomy), dividing label errors into localization errors and labeling errors, and further subdividing them into multiple sub - classes. - Based on this classification, the researchers developed a simple and effective heuristic method to automatically detect and identify incorrect labels. 4. **Experimental verification**: - The researchers conducted experiments on multiple commonly - used HPE datasets (such as COCO and MPII), verified the effectiveness of the proposed heuristic method, and demonstrated the improvement of model performance after data cleaning. - The experimental results show that removing incorrect labels can significantly improve the evaluation metrics of the HPE model, especially with obvious improvements in the detection accuracy of key parts such as hips, knees, and ankles. ### Summary The core problem of this paper is: **How to identify and deal with label errors in the human pose estimation dataset to improve the training quality and generalization ability of the model**. Through systematic analysis and experimental verification, the researchers have proven the critical impact of data quality on the performance of the HPE model and provided an effective method to meet this challenge.

The Influence of Faulty Labels in Data Sets on Human Pose Estimation

Self-Supervised Learning for Human Pose Estimation in Sports

ARHPE: Asymmetric Relation-aware Representation Learning for Head Pose Estimation in Industrial Human-machine Interaction

Quantifying and mitigating the impact of label errors on model disparity metrics

Investigating Labeler Bias in Face Annotation for Machine Learning

Pose Impact Estimation on Face Recognition using 3D-Aware Synthetic Data with Application to Quality Assessment

Source-free Domain Adaptive Human Pose Estimation

A Latent Clothing Attribute Approach for Human Pose Estimation

3D Human Pose Estimation Based on Wearable IMUs and Multiple Camera Views

Optimising 2D Pose Representation: Improve Accuracy, Stability and Generalisability Within Unsupervised 2D-3D Human Pose Estimation

Semi-Supervised 2D Human Pose Estimation Driven by Position Inconsistency Pseudo Label Correction Module

The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation

Efficient, Self-Supervised Human Pose Estimation with Inductive Prior Tuning

A Review on Human Pose Estimation

Design Space Exploration on Efficient and Accurate Human Pose Estimation from Sparse IMU-Sensing

Learning under Label Noise through Few-Shot Human-in-the-Loop Refinement

Domain Knowledge-Informed Self-Supervised Representations for Workout Form Assessment

Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats

Personalized 3D Human Pose and Shape Refinement

Don't Blame the Data, Blame the Model: Understanding Noise and Bias When Learning from Subjective Annotations

Generalized Pose Space Embeddings for Training In-the-Wild using Anaylis-by-Synthesis