Compositional Video Prediction

Summary

• Method for solving pixel-level future prediction given a single image
• Decompose scene into distinct entities that undergo motion and possibly interact
• Generates realistic predictions for stacked objects and human activities in gyms
• Links: [ website ] [ pdf ]

Background

• Given a single image, humans can easily understand the scene and make predictions
• Many other works also model relationships between objects, but do some combination of:
  • Use only simple visual stimuli
  • Use state based input
  • Rely on sequence of inputs
  • Don’t make predictions in pixel space

Methods

• Given starting frame $f^0$ and the location of $N$ entities $\{b_n^0{\}}_{n=1}^N$, predict $T$ future frames $f^1,f^2,\dots ,f^T$
• Use entity predictor $\mathcal{P}$ for per-entity representations $\{x_n^t{\}}_{n=1}^N$
  • $\{x_n^{t+1}\}\equiv \mathcal{P}(\{x_n^t\},z_t)$
  • $\{x_n^t{\}}_{n=1}^N\equiv \{(b_n^t,a_n^t){\}}_{n=1}^N$
  • $b_n^t$ is the predicted location
  • $a_n^t$ is the predicted implicit features for entity appearance
    • $a_n^0$ obtained using ResNet-18 on cropped region from $f^0$
  • Interaction between entities modeled with graph neural network, edges are predetermined
• Use frame decoder $\mathcal{D}$ to infer pixels
  • $f^t\equiv \mathcal{D}(\{x_n^t\},f^0)$
  • Warp normalized spatial representation to image coordinates with predicted location for each entity
  • Use soft masking for each entity to account for possible occlusions
  • Add these masked features to features from initial frame $f^0$
• Single random latent variable $u$ to capture multi-modality of prediction task
  • Yields per-timestep latent variables $z_t$, which are correlated accross time, via a learned LSTM
• Loss consists of three terms:
  • Prediction: $l_1$ loss on decoded frame from predicted features and $l_2$ loss on predicted location
  • Encoder: information bottleneck on latent variable distribution
  • Decoder: $l_1$ loss on decoded frame from features extracted from the same frame (auto-encoding loss)

Results

• Datasets:
  • ShapeStacks: synthetic dataset of stacked objects that fall, different block shapes/colors and configurations
  • Penn Action: real video dataset of people playing indoor/outdoor sports, annotated with joint locations
    • Use gym activities subset: less camera motion, more similar backgrounds
• Metrics:
  • Average MSE for entity locations
  • Learned Perceptual Image Patch Similarity (LPIPS) for generated frames
  • Due to stochasticity from random variable $u$, use best scores from 100 samples
• Baselines:
  • No-Factor: only predicts at the level of frames
  • No-Edge: no interaction among entities
  • Pose Knows: also predicts poses as intermediate representation, predicts location but not appearance, no interactions
• Proposed method consistenly performs better than the simple baselines on ShapeStacks, but No-Edge is a fairly close second
• Adding adversarial loss, from Pose Knows, greatly improves performance of proposed method on Penn Action
• Qualitative results on Penn Action are generally not that convincing for any model

Conclusion

• Relies on supervision of entity locations and graph structure
• Inclusion of latent variable to model multi-modality is interesting
  • Better metrics for evaluating diversity and accuracy would be useful
• Poor performance on real world dataset suggests there is still a lot of room for improvement