Jingdong Wang comments

Results 10 comments of


                                            Jingdong Wang

The HRNet+OCR branch

The performance on LIP and Pascal Context (COCO stuff and ADE) is stable. But on Cityscapes, the performance is not very stable.

The performance gap

@MandyMo: have you re-produced the results?

The performance gap

@MandyMo did you use the same settings? what are your results?

Inference time is suprisingly long

This is a good point. PyTorch does not support multi-branch structure well. The inference time is a little long. With careful implementation at CPU, the runtime acceleration is also the...

在COCO数据集上的人体姿态估计提供的detector是用的什么检测模型.

It is the same as https://github.com/leoxiaobin/deep-high-resolution-net.pytorch.

how to show pose keypoint

Visualization codes are available at https://github.com/leoxiaobin/deep-high-resolution-net.pytorch

miss "NUM_OUTS" in NECK config

@zimenglan-sysu-512 @haochange HRNets: The runtime computation cost in Pytorch is higher because multi-branch convolutions are run sequentially. Instead, in tensorflow, the runtime cost is much lower.

About the muli-scale testing on MPII test set, i only got 91.6% instead of 92.3% reported in the paper.

Thanks a lot for your interests! Please check our new version (simpler) https://github.com/HRNet/DEKR

Train.py

The reproduction of OCR with HRNet is available at https://github.com/HRNet/HRNet-Semantic-Segmentation/tree/HRNet-OCR.