1910.09070.md - hassony2/inria-research-wiki GitHub Wiki

Structured Prediction Helps 3D Human Motion Modelling, ICCV'19 {project page} {paper} {code} {notes}

Emre Aksan, Manuel Kaufmann, Otmar Hilliges

Objective

Generate good short-term (typically 400 ms) motion predictions. Reflect on correlation between quantitative and qualitative performance for this task. Typically, Euclidean distance on the Euler angles is used for evaluation.

Explicitely leverage kinematic constraints, using a SPL (Structured Prediction Layer).

Datasets

Human 3.6, saturated

AMASS dataset (14x larger and more diverse then H3.6)

Method

• SPL layer
  • each joint receives information from the context encoding and the parent joint’s prediction
  • can be seen as as a dense layer with some connections set to zero by leveraging domain knowledge (e.g. position of right arm has not a lot of influence on position of right leg)
  • each joint is modelled with only a small hidden layer + ReLU and linear projection to the joint prediction

Experiments

• Different hierarirchical structures (all parents in kinematic chain vs only direct parents)

• Improvement from SPL

  • very limited on top of QuaterNet, but more significant for seq2seq and RNN baselines
  • looks like most improvement on AMASS comes from per-joint loss and on H3.6 from the explicit kinematic chain constraints