Globally-Robust Instance Identification and Locally-Accurate Keypoint Alignment for Multi-Person Pose Estimation
Fangzheng Tian, Sungchan Kim
Abstract
Scenes with a large number of human instances are characterized by significant overlap of the instances with similar appearance, occlusion, and scale variation. We propose GRAPE, a novel method that leverages both Globally Robust human instance identification and locally Accurate keypoint alignment for 2D Pose Estimation. GRAPE predicts instance center and keypoint heatmaps, as global identifications of instance location and scale, and keypoint offset vectors from instance centers, as representations of accurate local keypoint positions. We use Transformer to jointly learn the global and local contexts, which allows us to robustly detect instance centers even in difficult cases such as crowded scenes, and align instance offset vectors with relevant keypoint heatmaps, resulting in refined final poses. GRAPE also predicts keypoint visibility, which is crucial for estimating centers of partially visible instances in crowded scenes. We demonstrate that GRAPE achieves state-of-the-art performance on the CrowdPose, OCHuman, and COCO datasets. The benefit of GRAPE is more apparent on crowded scenes (CrowdPose and OCHuman), where our model significantly outperforms previous methods, especially on hard examples.
It will take some time to load the videos and images when you open the webpage, thank you for your patience.
×
❮
❯
Qualitative Results
For easy viewing of details. Click on the image in this section to make it full screen. In full-screen mode, you can use the mouse wheel to zoom in on the image and drag the mouse to view the details. Press 'Esc' to exit full screen mode.
Figure M1 and Figure M2 are from Figure 5 and Figure 6 in the manuscript. We paste the relevant content as follows.
We provide examples of qualitative evaluation to demonstrate the advantages of our model over recent work by HigherHRNet [3], DEKR [1], and LOGO-CAP [2]. Figure M1 shows that DEKR [9] misses the keypoints of the people due to the low contrast of the image (1st and 2nd examples) or results in incorrect predictions due to the significantly overlapping human bodies (3rd and 4th examples), where GRAPE performs successfully. As shown in Figure M2, existing methods, HigherHRNet [3], DEKR [1], and LOGO-CAP [2], suffer from difficult situations, such as small instances and occlusion due to poor image contrast. While HigherHRNet [3] detects more instances than DEKR [1] and LOGOCAP [2], it performs poorly with instances of similar appearance, leading to unrealistic keypoint groupings (1st example). LOGOCAP [2] misses the most instances while it accurately predicts keypoints for detected instances. In contrast, our method mitigates this difficulty by robustly estimating the centers and scales of human instances, resulting in most instance detection and accurate keypoint prediction.
[1/7]
![Image 1](images/fig5.png)
Figure M1: Qualitative results of (a) GRAPE and (b) DEKR [1] on CrowdPose test.
[2/7]
![Image 2](images/fig6.png)
Figure M2: Qualitative results of GRAPE, HigherHRNet [3], DEKR [1], and LOGO-CAP [2] (from top to bottom) on COCO val OCHuman test. The OChuman results were obtained from the models trained on COCO.
We provide more qualitative evaluation examples on the three datasets (Figures S1, S2 and S3). We generate qualitative results based on publicly available implementations of HigherHRNet [3], DEKR [1], and LOGOCAP [2]. We set all thresholds used in Section 3.3 of the manuscript to 0.2.
Figure S1 shows comparisons of GRAPE with DEKR [1] on the CrowdPose test by training the models on CrowdPose train and val. DEKR [1] misses many keypoints, leading to false matches, where GRAPE effectively avoids these problems.
We show additional results of GRAPE by comparing with HigherHRNet [3], DEKR [1] and LOGO-CAP [2] on the COCO val and OChuman test in Figures S2 and S3, respectively.
Figure S4 provides failure cases with our model compared to previous methods.
We visualize the results of human keypoints predicted from each of the steps in our method in Figure S5.
[3/7]
![Image 2](images/fig_supp1.png)
Figure S1: Qualitative results on CrowdPose test: (a) GRAPE (proposed) and (b) DEKR [1]. Comparing predictions for small sized persons are highlighted in red.
[4/7]
![Image 2](images/fig_supp2.png)
Figure S2: Qualitative results on COCO val: GRAPE (proposed), HigherHRNet [2], DEKR [1], and LOGO-CAP [2].
[5/7]
![Image 2](images/fig_supp3.png)
Figure S3: Qualitative results on OCHuman test: GRAPE (proposed), HigherHRNet [2], DEKR [1], and LOGO-CAP [2]. All models were were trained on COCO.
[6/7]
![Image 2](images/fig_supp4.png)
Figure S4: Examples of failures with our model on CrowdPose test (1st and 2nd rows), COCO val (3rd and 4th rows), and OCHuman test (the last two rows). Corresponding results with other methods are also provided.
[7/7]
![Image 2](images/fig_supp5.png)
Figure S5: Qualitative results: the offset vectors predicted by (a) the backbone and (b) the Transformer, (c) the final keypoint predictions aligned with keypoint heatmaps where all keypoints are shown regardless of its visibility, and (d) overlaid keypoint predictions from (a), (b), and (c). Examples are taken from CrowdPose test (1st and 2nd rows), COCO val (3rd and 4th rows), and OCHuman test (the last two rows).
References
[1] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. 2021.Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14676–14686.
[2] Nan Xue, Tianfu Wu, Gui-Song Xia, and Liangpei Zhang. 2022. Learning LocalGlobal Contextual Adaptation for Multi-Person Pose Estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13065–13074.
[3] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. 2020. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5386–5395.