2D Human Pose Estimation: A Survey

1 Introduction

• Deep learning learns feature representations from data, advancing pose estimation beyond earlier methods that relied on hand-crafted features.The earlier probabilistic graphical-model methods were limited in generalization and performance.Source: §1 · p. 1
• Few earlier surveys focus on 2D human pose estimation, and none, to the authors’ knowledge, summarizes signal-based methods such as RF and radar.The authors also argue that dividing methods into single-person and multi-person categories may be unnecessary because a model can accommodate both scenes.Source: §1 · p. 2
• The survey groups methods into network architecture design, network training refinement, and model-agnostic post processing, and also discusses pose reconstruction from RF and radar signals.The categories cover feature extraction, training techniques, and polishing coarse pose estimates, respectively.Source: §1 · p. 2

1.1 Scope

• The review covers 2D human pose estimation with deep learning, excluding conventional non-deep-learning methods.Applications and body-model representations adequately covered by other reviews are not treated in detail.Source: §1.1 · p. 2
• To narrow the literature, the authors select top journal and conference papers published since 2014.Source: §1.1 · p. 2

1.2 Outline

• The paper proceeds from problem formulation and architecture design to training refinement, post processing, benchmarks, discussions, and conclusions.The discussions include open questions, signal-based 2D pose estimation, and future directions.Source: §1.2 · p. 2

2.1 The Problem

• 2D human pose estimation detects the poses of all people in an image or video, representing the output as P = {P_i} for n people.Source: §2.1 · p. 2
• Skeleton-based models represent the body as predefined joints and are widely used in 2D pose estimation; contour- and volume-based models are alternatives.Contour representations include rough body contour and limb width, while volume representations describe 3D human shapes.Source: §2.1 · p. 2

2.2 Technical Challenges

• High accuracy and efficiency are both desired, but accurate models can require more computation and storage, making real-time use difficult.HRNet-W48 is cited as a high-performing model that has difficulty achieving real-time pose estimation even on an NVIDIA GTX-1080TI; lightweight models with comparable precision are sought for mobile or wearable devices.Source: §2.2 · pp. 2–4
• The figure illustrates pose estimation from images, videos, and RF signals, including a skeleton extracted from RF alone.The RGB image accompanying the RF example is for visual reference.Source: Figure 1 · p. 3
• Accurate pose detection is challenged by exposure variation, person-object entanglement, occlusion, self-occlusion, motion blur, and video defocus.These conditions can lead to detection failure or confuse visual-feature-based keypoint detectors.Source: §2.2 · p. 3

3 Network Architecture Design Methods

• Network architecture design investigates model structures to extract powerful features for human pose estimation.The section organizes these approaches into top-down and bottom-up frameworks.Source: §3 · p. 4

3.1.1 Regression-Based Methods

• Regression-based methods directly map images to predefined joint coordinates, offering efficiency but not representing each body part’s spatial area.The survey contrasts direct coordinates with probabilistic heatmaps.Source: §3.1.1 · p. 4

3.1.2 Heatmap-Based Methods

• Heatmap-based methods encode each joint with a probability map; they are more popular for higher accuracy, but incur computational overhead and quantization error.Heatmaps use 2D Gaussians centered at joint locations.Source: §3.1.2 · pp. 4, 7

3.1.3 Video-Based Methods

• Video methods exploit temporal information through optical flow, recurrent networks, pose tracking, and key-frame approaches.Optical flow uses frame motion, while recurrent networks model temporal context and tracking builds person-specific tracklets.Source: §3.1.3 · pp. 7–8
• DCPose uses neighboring-frame information and is reported to achieve state-of-the-art performance on video-based human pose estimation.Its dual-direction framework processes consecutive frames with temporal merging, residual fusion, and pose correction.Source: §3.1.3 · pp. 8–9

3.1.4 Model Compression-Based Methods

• Compression methods target models that are resource-intensive and often not real-time, reducing parameters while aiming for only a small accuracy decline.Examples include teacher-student distillation and Lite-HRNet designs that simplify HRNet.Source: §3.1.4 · p. 9

3.1.5 Summary of Top-Down Framework

• Top-down systems combine a human detector that produces bounding boxes with a pose estimator; advances in either component can improve the scalable framework.The detector affects proposal quality, while the pose detector directly determines pose accuracy.Source: §3.1.5 · p. 9

3.2 Bottom-Up Framework

• Bottom-up methods skip human detection, localize identity-free keypoints, then group them using center regression, associate embeddings, or part fields.Eliminating object detection reduces computational overhead; part-field methods connect detected keypoints before grouping them.Source: §3.2 · pp. 9–10

4.1 Data Augmentation Techniques

• Data augmentation increases sample number and diversity to address the time and expense of building large, high-quality datasets.The surveyed techniques include random transformations and learned methods that generate difficult or synthetic poses, alter semantic body parts, or mix corrupted images.Source: §4.1 · p. 10

4.2 Multi-Task Training Strategies

• Multi-task methods share representations across related visual tasks, including human parsing, to improve pose estimation.Some jointly predict parsing and pose and use part-level segments to guide keypoint localization; another approach uses a parsing encoder and pose-model parameter adapter to learn complementary features.Source: §4.2 · p. 10

4.3 Loss Function Constraints

• Loss-function approaches modify the learning objective, from standard L2 heatmap error to structural, hard-keypoint, and distillation losses.The surveyed examples include multi-scale human-structure constraints, extra penalties for the top-M hard keypoints, and a combined distillation loss addressing structure, similar-joint confusion, and heatmap distributions.Source: §4.3 · pp. 10–11

4.4 Domain Adaption Methods

• Domain adaptation methods use labeled source data to adapt pose estimators to unlabeled or sparsely labeled target domains.Examples align body topology and fine-grained features, train across multiple datasets, or update pseudo-labels to reduce the synthetic-to-real gap; the latter is also reported for animal pose estimation.Source: §4.4 · p. 11

5.2 Pose Resampling

• Pose resampling refines initial estimates using additional pose cues, and model-agnostic methods can combine image context with pose structure.Examples include refining an input pose, estimating invisible joints after localizing visible ones, and graph-based keypoint refinement.Source: §5 · p. 11, §5.2 · p. 12

5.1 Quantization Error

• Heatmap-based keypoint decoding can incur quantization error, motivating methods that improve coordinate extraction or make heatmap-to-coordinate conversion differentiable.Distribution-aware modulation adjusts predicted heatmaps before decoding, while other approaches use unit-length data processing or differentiable Softargmax and integral methods.Source: §5.1 · pp. 11–12

5.2 Pose Resampling

• Video pose resampling aggregates pose estimates across frames, either treating predictions equally or weighting them using pose similarity.Other surveyed strategies cluster hypotheses and choose keypoints near cluster centers, or decode current-frame heatmaps from aggregated features.Source: §5.2 · p. 12

6.1 Benchmark Datasets

• The surveyed benchmarks span image and video pose estimation, with COCO described as the most popular image-based benchmark and PoseTrack covering articulated tracking in crowded scenes.COCO includes 250,000 pose annotations; PoseTrack2018 contains 1,138 clips and 153,615 pose annotations.Source: §6.1 · pp. 12–13, Table 2 · p. 13

6.2 Evaluation Metrics

• PCP and PCK assess localized body parts or keypoints against thresholds, while AP is based on Object Keypoint Similarity and is averaged differently for image and video estimation.PCK thresholds may use bounding-box size or head-segment length; image-based mAP averages AP across OKS thresholds.Source: §6.2 · p. 14

6.3 Performance Comparisons

• Top-down approaches outperform bottom-up approaches by 5.4 mAP on COCO; UDP with HRNet achieves state-of-the-art results without extra training data.The comparison concerns image-level pose estimation on COCO test-dev2017.Source: §6.3 · p. 14, Table 3 · p. 15
• Lite-HRNet offers an accuracy–speed trade-off, achieving 69.7 mAP with 1.8 M parameters.The survey identifies it among approaches designed as small networks.Source: §6.3 · p. 15
• On PoseTrack2017, DCPose uses temporal information from adjacent frames and establishes new state-of-the-art results on both validation and test sets.Source: §6.3 · p. 15

7.1 Open Questions

• Occlusion, person entanglement, and motion blur challenge pose localization; video datasets also have limited quantity, scene variety, and position labels for occluded joints.The discussion also identifies a need for domain-specific datasets, such as for dancing and swimming.Source: §7.1 · p. 16

7.2 Signal-Based Human Pose Estimation

• Radio signals and radar offer alternative inputs for pose estimation, including WiFi methods intended to estimate poses when a person is occluded by a wall.A cited 77GHz millimeter-wave radar system converts radar-generated heatmaps into poses with a CNN.Source: §7.2 · pp. 16–17

7.3 Future Directions

• Unsupervised learning is proposed as a direction because full annotations cannot be provided for the vast supply of unlabeled images and videos.Source: §7.3 · p. 17
• Heatmap encoding and decoding introduce quantization error and resolution-dependent computation, motivating a novel unbiased pose representation.Source: §7.3 · p. 17
• The authors call for more interpretability research because the effect of image features on final keypoint localization remains unclear.Source: §7.3 · p. 17

• The survey classifies methods as network architecture design, network training refinement, or post processing, then reviews benchmarks, metrics, performance comparisons, and future directions.Its finer categories include top-down and bottom-up frameworks; data augmentation, multi-task learning, loss constraints, and domain adaptation; and quantization error and pose resampling.Source: §8 · p. 17