摘要
近年来，姿态估计领域取得了显著进展，同时对姿态跟踪的兴趣也在不断增加。然而，随着算法和系统复杂性的整体提升，算法的分析与比较变得更加困难。本研究提供了简单且有效的基线方法，这些方法有助于激发新思路并评估该领域的创新想法。在具有挑战性的基准测试中，这些方法达到了最先进的结果。代码将在 https://github.com/leoxiaobin/pose.pytorch 上提供。
代码仓库
leeyegy/simcc
pytorch
GitHub 中提及
open-mmlab/mmpose
pytorch
leeyegy/SimDR
pytorch
GitHub 中提及
SunJuil-Ty/simplebaselines-mindspore
mindspore
lyqcom/alphapose
mindspore
GitHub 中提及
Microsoft/human-pose-estimation.pytorch
pytorch
GitHub 中提及
CuberrChen/HumanPoseBL-Paddle
paddle
GitHub 中提及
simochen/flowtrack.pytorch
pytorch
GitHub 中提及
mks0601/TF-SimpleHumanPose
tf
GitHub 中提及
osmr/imgclsmob
mxnet
GitHub 中提及
Mind23-2/MindCode-77
mindspore
GitHub 中提及
victimsnino/pose-simple-baselines-demo.pytorch.
pytorch
GitHub 中提及
dog-qiuqiu/Ultralight-SimplePose
mxnet
GitHub 中提及
dog-qiuqiu/MobileNetV2-SimplePose
mxnet
GitHub 中提及
mindspore-lab/mindone
mindspore
GitHub 中提及
leoxiaobin/pose.pytorch
官方
pytorch
bearpaw/pytorch-pose
pytorch
GitHub 中提及
samson6460/tf2_pose_estimation
tf
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 2d-human-pose-estimation-on-jhmdb-2d-poses | SimplePose | PCK: 94.4 |
| 2d-human-pose-estimation-on-ochuman | ResNet-152 | Test AP: 33.3 Validation AP: 41.0 |
| 2d-human-pose-estimation-on-ochuman | ResNet-50 | Test AP: 30.4 Validation AP: 37.8 |
| keypoint-detection-on-coco | ResNet-50 | Validation AP: 72.2 |
| keypoint-detection-on-coco-test-challenge | Simple Base+* | AP: 74.5 AP50: 90.9 AP75: 80.8 APL: 87.5 AR: 80.5 AR50: 95.1 AR75: 86.3 ARL: 82.9 ARM: 75.3 |
| keypoint-detection-on-coco-test-dev | Simple Base | AP50: 91.9 AP75: 81.1 APL: 80.0 APM: 70.3 AR: 79.0 |
| keypoint-detection-on-coco-test-dev | Simple Base+* | AP50: 92.4 AP75: 84.0 APL: 82.7 APM: 73.0 AR: 81.5 AR50: 95.8 AR75: 88.2 ARL: 87.2 ARM: 77.4 |
| keypoint-detection-on-ochuman | ResNet-50 | Test AP: 29.5 Validation AP: 32.1 |
| keypoint-detection-on-ochuman | ResNet-152 | Test AP: 33.3 Validation AP: 41.0 |
| multi-person-pose-estimation-on-crowdpose | Simple baseline | AP Easy: 71.4 AP Hard: 51.2 AP Medium: 61.2 mAP @0.5:0.95: 60.8 |
| multi-person-pose-estimation-on-ochuman | SimplePose | AP50: 37.4 AP75: 26.8 Validation AP: 24.1 |
| pose-estimation-on-aic | SimpleBaseline (ResNet-152) | AP: 29.9 AP50: 73.8 AP75: 18.3 AR: 34.3 AR50: 76.9 |
| pose-estimation-on-aic | SimpleBaseline (ResNet-101) | AP: 29.4 AP50: 73.6 AP75: 17.4 AR: 33.7 AR50: 76.3 |
| pose-estimation-on-aic | SimpleBaseline (ResNet-50) | AP: 28.0 AP50: 71.6 AP75: 15.8 AR: 32.1 AR50: 74.1 |
| pose-estimation-on-coco-test-dev | Flow-based (ResNet-152) | AP: 73.7 AP50: 91.9 AP75: 81.1 APL: 80 APM: 70.3 AR: 79 |
| pose-estimation-on-coco-val2017 | SimpleBaseLine (256x192) | AP: 70.4 AP50: - AP75: - AR: - |
| pose-estimation-on-ochuman | ResNet-152 | Test AP: 33.3 Validation AP: 41.0 |
| pose-estimation-on-ochuman | ResNet-50 | Test AP: 29.5 Validation AP: 32.1 |
| pose-tracking-on-posetrack2017 | MSRA (FlowTrack) | MOTA: 57.81 mAP: 74.57 |
| pose-tracking-on-posetrack2018 | MSRA | MOTA: 61.37 mAP: 74.03 |