Predicting and understanding the dynamic of human motion has many applications such as motion synthesis, augmented reality, security, education, reinforcement learning, autonomous vehicles, and many others. In this thesis, we create a novel end-to-end pipeline that can predict multiple future poses from the same input, and, in addition, can classify the entire sequence. Our focus is on the following two aspects of human motion understanding:
Probabilistic human action prediction: Given a sequence of human poses as input, we sample multiple possible future poses from the same input sequence using a new GAN-based network.
Human motion understanding: Given a sequence of human poses as input, we classify the actual action performed in the sequence and improve the classification performance using the presentation learned from the prediction network.
We also demonstrate how to improve model training from noisy labels, using facial expression recognition as an example. More specifically, we have 10 taggers to label each input image, and compare four different approaches: majority voting, multi-label learning, probabilistic label drawing, and cross-entropy loss. We show that the traditional majority voting scheme does not perform as well as the last two approaches that fully leverage the label distribution. We shared the enhanced FER+ data set with multiple labels for each face image with the research community (https://github.com/Microsoft/FERPlus).
For predicting and understanding of human motion, we propose a novel sequence-to-sequence model trained with an improved version of generative adversarial networks (GAN). Our model, which we call HP-GAN2, learns a probability density function of future human poses conditioned on previous poses. It predicts multiple sequences of possible future human poses, each from the same input sequence but seeded with a different vector z drawn from a random distribution. Moreover, to quantify the quality of the non-deterministic predictions, we simultaneously train a motion-quality-assessment model that learns the probability that a given skeleton pose sequence is a real or fake human motion.
In order to classify the action performed in a video clip, we took two approaches. In the first approach, we train on a sequence of skeleton poses from scratch using random parameters initialization with the same network architecture used in the discriminator of the HP-GAN2 model. For the second approach, we use the discriminator of the HP-GAN2 network, extend it with an action classification branch, and fine tune the end-to-end model on the classification tasks, since the discriminator in HP-GAN2 learned to differentiate between fake and real human motion. So, our hypothesis is that if the discriminator network can differentiate between synthetic and real skeleton poses, then it also has learned some of the dynamics of a real human motion, and that those dynamics are useful in classification as well. We will show through multiple experiments that that is indeed the case.
Therefore, our model learns to predict multiple future sequences of human poses from the same input sequence. We also show that the discriminator learns a general representation of human motion by using the learned features in an action recognition task. And we train a motion-quality-assessment network that measure the probability of a given sequence of poses are valid human poses or not.
We test our model on two of the largest human pose datasets: NTURGB-D, and Human3.6M. We train on both single and multiple action types. The predictive power of our model for motion estimation is demonstrated by generating multiple plausible futures from the same input and showing the effect of each of the several loss functions in the ablation study. We also show the advantage of switching to GAN from WGAN-GP, which we used in our previous work. Furthermore, we show that it takes less than half the number of epochs to train an activity recognition network by using the features learned from the discriminator.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/d8-sq89-mm29 |
| Date | January 2019 |
| Creators | Barsoum, Emad |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0024 seconds