search

microsoft / human-pose-estimation.pytorch

The project is an official implement of our ECCV2018 paper "Simple Baselines for Human Pose Estimation and Tracking(https://arxiv.org/abs/1804.06208)"
MIT License
2.93k stars 603 forks source link

poor performance of mobilenetv1,v2 and mnasnet as backbone ??? #84

Open hengck23 opened 5 years ago

hengck23 commented 5 years ago

I experiment your framework with different backbone. Surprisingly, the performance of resnet34 is quite close to resnet50. Renest34 have mAP of about 0.705 (and 0.893 AP@0.5) on 2017 coco keypoints for 256x192, tested using ground truth box as input and without flip.

However, for mobilenetv1,v2 and mnasnet, their results are very bad, even worse then resnet18. Do you have any idea why this is so?

mobilenetv1,v2 and mnasnet has only mAP of about 0.648 on coco keypoint. Resnet18 is about 0. 657.

On imagenet, mobilenetv2(top-1 28%) and mnasnet (top-1 26%) has performance much better than resnet18 (top-1 30%), and is comparable to Resnet 34(top-1 27%) . Resnet50 is top-1 24%

PaTricksStar commented 5 years ago

I tried moilenet v2 and AP is 0.625. Thought it should be comparable to resnet50.

TTMRonald commented 5 years ago

@hengck23 @PaTricksStar I tried to train moilenet v3, the AP is 0.686. But GPU FPS only is 2. I want to know how much the moilenet v2 FPS is. thanks!

vivek87799 commented 4 years ago

@hengck23 Can you kindly help me with resnet 18 pre-trained weights.

tucachmo2202 commented 3 years ago

@hengck23 how do you train with other backbone? would you release your code in github or could you please give me the pretrained of those model? Thanks a lot!