Giters

HRNet / HRNet-Bottom-Up-Pose-Estimation

This is an official pytorch implementation of “Bottom-Up Human Pose Estimation by Ranking Heatmap-Guided Adaptive Keypoint Estimates” (https://arxiv.org/abs/2006.15480).

Home Page:https://github.com/HRNet/HRNet-Bottom-Up-Pose-Estimation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Training problem

jinleibupt opened this issue · comments

commented

I retrain your model, but I found the model can't get 67.3 with single scale. (ps. I only get 60.9 without the rescore model) Can you present more training detials? Thanks!

commented

We trained the model using 4 NVIDIA V100 GPU and this config file (experiments/coco/w32/w32_4x_reg_delaysep_bg01_stn_512_adam_lr1e-3_coco_x140.yaml) to get 67.3AP in COCO dataset. I would like to ask what changes have you made to this file?

commented

I did't change this file. However, I used 8 gpu. I found there is no val process in your training code. (ps. I add a val process in my code which can improve the result, but it still can't get 67.3AP)

commented

You can use tools/valid.py to evaluate your model. I think the performance is influenced by the post-processing.