# 2022/10/04 Meeting
###### tags: `Meeting`

## 3D human pose estimation

### One Stage Approaches:
* Directly regresses the 3D pose without the need to go through a 2D pose estimator.
* Usually, techniques that use ConvNets to directly regress the 3D pose involve a kind of ignorance of the dependencies between human joints.
To address this issue, a deep learning framework for structured 3D human pose prediction from monocular images is proposed.
* Due to the difficulty of obtaining the 3D ground truth of the human pose in outdoor environments, most existing datasets are labeled in a laboratory environment using motion capture systems. Consequently,this can cause overfitting during training hence the variations in background, viewpoints, and lighting are limited in the indoor environment. In order to solve this problem, GANs are implemented to generate valid poses for indoors and out doors images.


### Two Stage Approaches:
Two stage approaches infer the 3D pose of the subject based on 2D joint locations, the basic architecture of these models implements cascaded 2D and 3D regressors.

As a solution to the lack of large-scale 3D pose annotation datasets, several works take advantage of the large-scale 2D datasets either by using a weakly-supervised paradigm or a self-supervised paradigm.

Dabral et al. [86] suggested two novel anatomically inspired loss functions, respectively named as illegal-angle and symmetry loss, and use them with a weakly supervised learning framework to train the model using both large-scale outdoor 2D data and limited indoor 3D data. The proposed loss function is applied to the 2D images during training and ensures that the predicted 3D pose does not violate anatomical constraints.

*此處提到的方法多為Multiview的method(Ex: EpipolarPose)

Takahashi et al. [95] proposed a new algorithm to estimate the
human 3D pose from multi-view videos captured by unsynchronized
and non-calibrated cameras. First, the network detects 2D joint positions from multi-view videos using a 2D pose detector


### Results and Comparison