gold:
learn a geometry-aware 3D representation G for the human pose
discover the geometry relation between paired images (Iti,Itj)
which are acquired from synchronized and calibrated cameras
- i,j:different view point
- t:acquiring time acquiring
main components
- image skeleton mapping
- skeleton-based view synthesis
- representation consistency constraint
image-skeleton mapping
input raw image pair (Iti,Itj) with size of W×H
- i,j:different view point ,(synchronized and calibrated cameras i,j)
Cti,Ctj:K keypoint heatmaps from a pre-trained 2D human pose eastimator
We follow previous works [45, 20, 18] to train the 2D estimator on MPII dataset.
constructed 8 pixels width 2D skeleton maps from heatmaps
binary skeleton maps pair (Sti,Stj),St(⋅)∈{0,1}(K−1)×W×H
Geometry representation via view synthesis
training set T={(Sti,Stj,Ri→j)}
- pairs of two views of projection of same 3D skeleton (Sti,Stj)
- relative rotation matrix Ri→j from coordinate system of camera i to j
Straightforward way for learning representation in unsupervised/weakly-supervised manner is to utilize auto-encoding mechanism reconstrcting input image.
- no geometry structure information
- nor provides more useful information for 3D pose estimation than 2D coordinates
novel ‘skeleton-based view synthesis’ generate image under a new viewpoint.
Given an image under the known viewpoint as input
source domain : (input image) Si={Sti}i=1V
- V :amount of viewpoints
target domain :generate image Sj={Stj}i=1V
- j̸=i
encoder ϕ:Si→G
- source skeleton Sti→Si into a latent space Gi∈G
decoder ψ:Ri→j×G→Si
- ratation matrix Ri→j
-
Gi as the set of m discrete points on 3η-dimensional space
in practice: G=[g1,g2,...gM]T,gm=(xm,ym,zm)
Lℓ2(ϕ⋅ψ,θ)=NT1∑∥ψ(Ri→j×ϕ(Sti))−Stj∥
Representation consistency constraint
image skeleton mapping + view synthesis (previous two steps)lead to unrealistic generation on target pose when there are large self occlusions in source view
Since there is no explicit constraint on latent space to facilitate G to be semantic
We assume there exists an inverse mapping (one-to-one) between source domain and target domain, on the condition of the known relative rotation matrix. We could find:
a encoder μ:Sj→G
- maps target skeleton Stj to latent space G~j∈G
a decoder ν:Rj→i×G→Sti
- maps representation G~j back to source skeleton Sti
- G~i and Gi should be the same shared representation on G with different rotation-related coefficients ------namely representation consistency
lrc=m=1∑M∥f×Gi−G~i∥2
- f=Ri→j rotation-related transformation that map Gi to G~j
- a bidirectional encoderdecoder framework
- generator(ϕ,ψ),Gij:rotated Gi
-
generator(μ,ν)
total loss of the bidirectional model
where θ and ζ denotes the parameters of two encode-decoder networks, respectively.
3D human pose estimation by learnt representation
given: monocular image I
goal : b={(xp,yp,zp)}p=1P,P body joints,b∈B
function F:I→B to learn the pose regression
2 fully connect layer