search

Daniil-Osokin / lightweight-human-pose-estimation.pytorch

Fast and accurate human pose estimation in PyTorch. Contains implementation of "Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose" paper.
Apache License 2.0
2.08k stars 473 forks source link

About net structure #35

Closed MoonBunnyZZZ closed 5 years ago

MoonBunnyZZZ commented 5 years ago

@Daniil-Osokin Q1: Does cpm in the network mean Convolutional Pose Machines? Q2: Where did cpm idea come from?It seems like that it's very important for result Q3: In #27, ELU was talked.Is it a good idea using ELU in InitialStage and RefinementStage? Because , in InitialStage, BN is not used.In RefinementStage,some layers have no BN after ReLU.

Daniil-Osokin commented 5 years ago

Hi, cpm means Convolutional Pose Machines, in the code this just the name for additional layers to adopt features of backbone for pose estimation. I believe using ELU is a good idea to tackle dead neurons problem, and as it was discussed in #27 in this code this is a legacy to reproduce the paper results. Absence of BN in some layers is also legacy, we've released new code for single-person pose estimation with more BNs.

MoonBunnyZZZ commented 5 years ago

Hi, cpm means Convolutional Pose Machines, in the code this just the name for additional layers to adopt features of backbone for pose estimation. I believe using ELU is a good idea to tackle dead neurons problem, and as it was discussed in #27 in this code this is a legacy to reproduce the paper results. Absence of BN in some layers is also legacy, we've released new code for single-person pose estimation with more BNs.

Thank you for your reply.I will enjoy your new work.

MoonBunnyZZZ commented 5 years ago

Hi, cpm means Convolutional Pose Machines, in the code this just the name for additional layers to adopt features of backbone for pose estimation. I believe using ELU is a good idea to tackle dead neurons problem, and as it was discussed in #27 in this code this is a legacy to reproduce the paper results. Absence of BN in some layers is also legacy, we've released new code for single-person pose estimation with more BNs.

Did you ever test ELU in gccpm-look-into-person-cvpr19.pytorch?

Daniil-Osokin commented 5 years ago

No, it was not the purpose there.

MoonBunnyZZZ commented 5 years ago

No, it was not the purpose there.

In your opinion, if using mobilenet v3 as backbone, which one layer to truncate？

Daniil-Osokin commented 5 years ago

By default, try to reuse all features, except classifier. Just add dilations and remove strides to save spatial resolution (have heatmaps and pafs 8 times smaller than input image).

Daniil-Osokin commented 5 years ago

Hope, it helped.