HRNet
inference time
Have u test, I tried it needs about 2 sec/pic.
@leoxiaobin have u test the time
I have implemented a simplified version of HigherHRNet (compatible with official pretrained weights) providing a live demo script: https://github.com/stefanopini/simple-HigherHRNet My implementation runs at about 3 fps on a GTX 1060 (single scale).
@hellojialee I'm sorry it doesn't, I still haven't implemented it. However, predictions on original and flipped images are combined before most of the post-processing so the inference time should be faster than 1.5fps.
@leoxiaobin @stefanopini @hellojialee have u test image using HigherHRNet，it seems that it‘s demo results is worser than centernet
@Dantju looking at the results reported in the papers, HigherHRNet should perform slightly better than CenterNet. However, the top-down HRNet outperforms HigherHRNet (see Table 2 in the paper) and in my experience it is quite faster.
CenterNet https://arxiv.org/pdf/1904.07850.pdf HigherHRNet https://openaccess.thecvf.com/content_CVPR_2020/papers/Cheng_HigherHRNet_Scale-Aware_Representation_Learning_for_Bottom-Up_Human_Pose_Estimation_CVPR_2020_paper.pdf
@stefanopini yes in paper HIgherHRNet is better than centernet,but i use centernet and higherHRNet to test same image,it seems that the demo result of centernet is better than HigherHRNet，I confused
