Recurrent Network Models for Human Dynamics

Katerina Fragkiadaki, Sergey Levine, Panna Felsen, Jitendra Malik

Seminal August 2015

TL;DR ERD wraps stacked LSTMs in learned nonlinear encoder and decoder networks to generate motion capture, label body joints in video and forecast them 400 ms ahead. On Human3.6M its generated walking stays realistic long after a plain LSTM collapses to the mean pose, and it fixes left-right joint confusions of a per-frame CNN detector.

What they do

Predicting human motion calls for a learned model of human dynamics. The Encoder-Recurrent-Decoder (ERD) puts nonlinear encoder and decoder networks before and after stacked LSTM layers and trains all three jointly, so the representation and its dynamics are learned together. The same design is applied on Human3.6M to three tasks: generating motion capture, labeling 2D body joints in video, and forecasting them 400 ms ahead.

How it works

Mocap generation. Fully connected encoder and decoder around the LSTMs; joint angles are exponential maps, with global position and heading relative to the previous frame. The output is one pose (Euclidean loss) or a 5-component Gaussian mixture over the next frame, and predictions are fed back as input to synthesize long sequences.
Denoising curriculum. Inputs are corrupted with Gaussian noise of increasing standard deviation, so the unrolled model learns to correct its own drift.
Video labeling and forecasting. A five-layer CNN encoder, initialized from a per-frame body part detector, passes CONV5 features of the person’s bounding box to the LSTM; a two-layer decoder outputs joint heat maps (6x6 and 12x12 scales combined) for the current or a future frame.
Replacing the encoder and decoder with linear maps underfit on all three tasks.

Results

Mocap prediction (Walking, test subject, 8 prefixes): a plain 3-layer LSTM has the lowest error at every horizon up to 560 ms (0.41 at 80 ms, 2.26 at 560 ms, against ERD’s 0.89 and 3.41; CRBM 3.34, N-gram 4.53, GPDM 4.61 at 560 ms). Qualitatively, that LSTM soon converges to the mean pose and the CRBM drifts to implausible motion, while ERD stays realistic for long periods.
Video labeling: ERD beats the per-frame CNN detector and a Viterbi smoothing baseline, mainly by fixing left-right confusions; Viterbi is only marginally better than the detector. Initializing the encoder from the detector is crucial.
Forecasting 400 ms ahead: ERD beats zero-motion and constant-optical-flow baselines, most clearly for the often-occluded lower limbs.

Limitations

ERD interpolates among training subjects rather than extrapolating, its short-term completions are less smooth, and it works better on periodic activities (walking, smoking) than non-periodic ones (sitting). It needs large training sets: on FlicMotion (170 training videos of about 50 frames) it does much worse than the per-frame detector and Viterbi. Mocap in-filling and denoising are not supported.

Figures

From Figure 1: ERD models for mocap generation (predicts the mocap vector at the next time step), video pose forecast... — From Figure 1: ERD models for mocap generation (predicts the mocap vector at the next time step), video pose forecasting (body joint heat maps of a future frame) and video pose labeling (heat maps of the current frame).

From Figure 2: Motion synthesis from a test-subject prefix (green), generated motion in blue. LSTM-3LR and CRBM give ... — From Figure 2: Motion synthesis from a test-subject prefix (green), generated motion in blue. LSTM-3LR and CRBM give smooth short-term completions, but LSTM-3LR soon converges to the mean pose and CRBM diverges to implausible motion, while ERD stays realistic for longer. NGRAM jumps at the transition and GPDM's temporal evolution is unrealistic.

From Figure 4: Video pose labeling on H3.6M for elbows, wrists, knees and ankles, accuracy versus normalized toleranc... — From Figure 4: Video pose labeling on H3.6M for elbows, wrists, knees and ankles, accuracy versus normalized tolerance radius. ERD beats the per-frame CNN detector (PF) and dynamic programming (VITERBI); the oracle curve is the upper bound set by the 12x12 heat map grid.

From Figure 5: ERD corrects left-right confusions of the per-frame CNN detector by aggregating CONV5 appearance featu... — From Figure 5: ERD corrects left-right confusions of the per-frame CNN detector by aggregating CONV5 appearance features over time.

Abstract

We propose the Encoder-Recurrent-Decoder (ERD) model for recognition and prediction of human body pose in videos and motion capture. The ERD model is a recurrent neural network that incorporates nonlinear encoder and decoder networks before and after recurrent layers. We test instantiations of ERD architectures in the tasks of motion capture (mocap) generation, body pose labeling and body pose forecasting in videos. Our model handles mocap training data across multiple subjects and activity domains, and synthesizes novel motions while avoid drifting for long periods of time. For human pose labeling, ERD outperforms a per frame body part detector by resolving left-right body part confusions. For video pose forecasting, ERD predicts body joint displacements across a temporal horizon of 400ms and outperforms a first order motion model based on optical flow. ERDs extend previous Long Short Term Memory (LSTM) models in the literature to jointly learn representations and their dynamics. Our experiments show such representation learning is crucial for both labeling and prediction in space-time. We find this is a distinguishing feature between the spatio-temporal visual domain in comparison to 1D text, speech or handwriting, where straightforward hard coded representations have shown excellent results when directly combined with recurrent units.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong