1 Introduction
Imaging watching a soccer game on television. At any given time, you can only see a subset of the players, and you may or may not be able to see the ball, yet you probably have some reasonable idea about where all the players currently are, even if they are not in the field of view. (For example, the goal keeper is probably close to the goal.) Similarly, you cannot see the future, but you may still be able to predict where the “agents” (players and ball) will be, at least approximately. Crucially, these problems are intertwined: we are able to predict future states by using a state dynamics model, but we can also use the same dynamics model to infer the current state of the world by extrapolating from the last time we saw each agent.
In this paper, we present a unified approach to state estimation and future forecasting for problems of this kind. More precisely, we assume the observed data consists of a sequence of video frames,
, obtained from a stationary or moving camera. The desired output is a (distribution over a) structured representation of the scene at each time step, , as well as a forecast into the future, , where encodes the state (e.g., location) of the ’th agent and .^{*}^{*}* In this work, we assume the number of agents is known, for simplicity — we leave the ”open world” scenario to future work.The classical approach to this problem (see, e.g., BarShalom et al. (2011)
) is to use statespace models, such as Kalman filters, for tracking and forecasting, combined with heuristics, such as nearest neighbor, to perform data association (i.e., inferring the mapping from observations to latent objects). Such generative approaches require a dynamical model for the states,
, and a likelihood model for the pixels, . These are then combined using Bayes’ rule. However, it is hard to learn good generative model of pixels, and inverting such models is even harder. By contrast, our approach is discriminative, and learns an inference network to compute the posterior belief statedirectly. In particular, our model combines ideas from graph networks, variational autoencoders, and RNNs in a novel way, to create what we call a graphstructured variational recurrent neural network (GraphVRNN).
We have tested our approach on two datasets: real basketball trajectories, rendered as a series of (partially observed) bird’s eye views of the court; and a simple simulated soccer game, rendered using a 3d graphics engine, and viewed from a simulated moving camera. We show that our approach can infer the current state more accurately than other methods, and can also make more accurate future forecasts. We also show that our method can vary its beliefs in a qualitatively sensible way. For example, it “knows” the location of the goalie even if it is not visible, since the goalie does not move much (in our simulation). Thus it learns to “see beyond the pixels”.
In summary, our main contribution is a unified way to do state estimation and future forecasting at the level of objects and relations directly from pixels using GraphVRNN. We believe our technique will have a variety of other applications beyond analysing sports videos, such as selfdriving cars (where inferring and predicting the motion of pedestrians and vehicles is crucial), and humanrobot interaction.
2 Related work
In this section, we briefly review some related work.
Graphstructured and stochastic RNNs.
There are several papers that combine standard RNNs (recurrent neural networks) with graphstructured representations, such as Chang et al. (2017); Battaglia et al. (2016); SanchezGonzalez et al. (2018). There are also several papers that extend standard RNNs by adding stochastic latent variables, notably the variational RNN approach of Chung et al. (2015), as well as recent extensions, such Krishnan et al. (2017); Fraccaro et al. (2016); Karl et al. (2017); Goyal et al. (2017); Buesing et al. (2018). However, as far as we know, combining VRNNs with graphs is novel.
Predicting future pixels from past pixels.
There are many papers that try to predict the future at the pixel level (see e.g., (Kitani et al., 2017) for a review). Some use a single static stochastic variable, as in a conditional VAE (Kingma & Welling, 2014), which is then “decoded” into a sequence using an RNN, either using a VAEstyle loss (Walker et al., 2016; Xue et al., 2016) or a GANstyle loss (Vondrick et al., 2016b; Mathieu et al., 2016). More recent work, based on VRNNs, uses temporal stochastic variables, e.g., the SV2P model of Babaeizadeh et al. (2018) and the SVGLP model of Denton & Fergus (2018). There are also various GANbased approaches, such as the SAVP approach of Lee et al. (2018) and the MoCoGAN approach of Tulyakov et al. (2018). The recent sequential AIR (attend, infer, repeat) model of Kosiorek et al. (2018)
uses a variablesized, objectoriented latent state space rather than a fixeddimensional unstructured latent vector. This is trained using a VAEstyle loss, where
is an RNN image generator.Forecasting future states from past states.
There are several papers that try to predict future states given past states. In many cases the interaction structure between the agents is assumed to be known (e.g., using fully connected graphs or using a sparse graph derived from spatial proximity), as in the social LSTM (Alahi et al., 2016) and the social GAN (Gupta et al., 2018), or methods based on inverse optimal control (Kitani et al., 2017; Lee et al., 2017). In the “neural relational inference” method of Kipf et al. (2018), they infer the interaction graph from trajectories, treating it as a static latent variable in a VAE framework. Ehrhardt et al. (2017)
proposed to predict a Gaussian distribution for the uncertain future states of a sphere. It is also possible to use graph attention networks
(Veličković et al., 2018) for this task (see e.g., Hoshen (2017); Battaglia et al. (2018)).Forecasting future states from past pixels.
Our main interest is predicting the hidden state of the world given noisy visual evidence. (Vondrick et al., 2016a) train a conditional CNN to predict future feature vectors, and several papers (e.g., (Lerer et al., 2016; Mottaghi et al., 2016; Wu et al., 2017; van Steenkiste et al., 2018; Fragkiadaki et al., 2016; Watters et al., 2017)) predict future object states. Our work differs by allowing the vision and dynamics model to share and update a common belief space; also, our model is stochastic, not deterministic. Our method is related to the “deep tracking” approach of (Dequaire et al., 2018). However, they assume the input is a partially observed binary occupancy grid , and the output is a fully observed binary occupancy grid, whereas we predict the state of individual objects using a graphstructured model. Our method is also related to the ”backprop Kalman filtering” approach of (Haarnoja et al., 2016). However, they assume the structure and dynamics of the latent state is known, and they use the Kalman filter equations to integrate information from the dynamical prior and (a nonlinear function of) the observations, whereas we learn the dynamics, and also learn how to integrate the two sources of information using attention.
3 Methods
3.1 VRNNs
We start by reviewing the VRNN approach of Chung et al. (2015). This model has three types of variable: the observed output , the stochastic VAE state , and the deterministic RNN hidden state , which summarizes and . We define the following update equations:
(1)  
(2)  
(3)  
(4) 
where are neural networks (see Section 3.4 for the details).
VRNNs are trained by maximizing the evidence lower bound (ELBO), which can be decomposed as follows:
(5) 
We use Gaussians for the prior and posterior, so we can leverage the reparameterization trick to optimize this objective using SGD, as explained in Kingma & Welling (2014). In practice, we scale the KL divergence term by , which we anneal from 0 to 1, as in (Bowman et al., 2016). Also, we start by using the ground truth values of (teacher forcing), and then we gradually switch to using samples from the model; this technique is known as scheduled sampling (Bengio et al., 2015), and helps the method converge.
3.2 Adding graph structure
To create a structured, stochastic temporal model, we associate one VRNN with each agent. The hidden states of the RNNs interact using a fully connected graph interaction network. (This ensures the model is permutation invariant to the ordering of the agents.) To predict the observable state for agent , we decode the hidden state vector using an MLP with shared parameters. The resulting model and operations are illustrated in Figure 2.
3.3 Conditioning on images
We can create a conditional generative model of the form by using a VRNN (where is the generated output) by making all the equations in Section 3.1 depend on as input. This is similar to a standard sequencetosequence model, but augmented to the graph setting, and with additional stochastic noise injected into the latent dynamics. We choose to only condition the outputs on the visual input , rather than making the hidden state, , depend on . The reason is that we want to be able to perform future prediction in the latent state space without having to condition on images. In addition, we want the latent noise to reflect variation of the dynamics, not the appearance of the world. See Figure 1
for a simplified illustration of our model (ignoring the graph network component). All parameters are shared across all agents except for the visual encoder; they have a shared feature extraction ”backbone”, but then use agentspecific features to identify the specific agent. (Agents have the same appearance across videos.) Thus the model learns to perform data association.
The key issue is how to define the output decoder, , in such a way that we properly combine the information from the current (partially observed) visual input, , with our past beliefs about the object states, , as well as any stochastic noise (used to capture any residual uncertainty). In a generative model, this step would be performed with Bayes’ rule, combining the dynamical prior with the visual likelihood. In our discriminative setting, we can learn how to weight the information sources using attention. In particular, we define
(6) 
where is the visible decoder, is the hidden decoder, and the terms are attention weights, computed as follows:
(7) 
where .
This is similar to the gating mechanism in an LSTM, where we either pass through the current observations, or we pass through the prior hidden state, or some combination of the two. If the visual input is uninformative (e.g., for future frames, we set ), the model will rely entirely on its dynamics model. To capture uncertainty more explicitly, we also pass in into the functions that compute attention, where is the set of “heatmaps” over object locations. (This is illustrated by the edges in Figure 1.) We also replace the sample with its sufficient statistics, , computed using the state dependent prior, .
To encourage the model to learn to forecast future states, in addition to predicting the current state, we modify the above loss function to maximize a lower bound
^{†}^{†}† Strictly speaking, the expression is only a lower bound if . See e.g., (Alemi et al., 2018) for a discussion of the KL weighting term. on , computed as follows:(8) 
where is a weighting factor that trades off state estimation and forecasting losses, is the prior, is the variational posterior, the expectation is over , and where we set if .
3.4 Implementation details
For the image encoder, we use convolutional neural networks with random initialization. In the case when there are multiple frames per step, we combine the image network with an S3D
(Xie et al., 2018) inspired network, which adds additional temporal convolutions on spatial feature maps. For the recurrent model, we tried both vanilla RNNs and GRUs, and found that results were similar. We report results using the GRU model. For the graph network, we use relation networks (Santoro et al., 2017), i.e. a fully connected topology with equal edge weights. During training, both state estimation and future prediction losses are important, and need to be weighted properly. When the future prediction loss weight is too small, the dynamics model is not trained well (reflected by loglikelihood). We find that scaling the losses using an exponential discount factor for future predictions is a good heuristic. We normalize the total weights for state estimation losses and future predictions losses to be the same.4 Results
In this section, we show results on two datasets, comparing our full model to various baselines.
4.1 Datasets
Since we are interested in inferring and forecasting the state of a system composed of multiple interacting objects, based on visual evidence, analysing sports videos is a natural choice. We focus on basketball and soccer, since these are popular games for which we can easily get data. The states here are the identities of the objects (players and ball), and their locations on the ground plane. Although specialized solutions to state estimation for basketball and soccer already exist (e.g., based on wearable sensors (Sanguesa, 2017), multiple calibrated cameras (Manafifard et al., 2017), or complex monocular vision pipelines (Rematas et al., 2018)), we are interested in seeing how far we can get with a pure learning based approach, since such a method will be more generally applicable.
Basketball.
We use the basketball data from Zhan et al. (2018). Each example includes the trajectories of 11 agents (5 offensive players, 5 defensive players and 1 ball) for 50 steps. We follow the standard setup from Zhan et al. (2018) and just model the 5 offensive players, ignoring the defense. However, we also consider the ball, since it has very different dynamics than the players. Overall, there are 107,146 training and 13,845 test examples. We generate bird’seye view images based on the trajectories (since we do not have access to the corresponding video footage of the game), where each agent is represented as a circle colorcoded by its identity. To simulate partial observation, we randomly remove one agent from the rendered image every 10 steps. Some example rendered images can be found in Figure 2(b).
Soccer.
To evaluate performance in a more visually challenging environment, we consider soccer, where it is much harder to get a bird’seye view of the field, due to its size. Instead most soccer videos are recorded from a moving camera which shows a partial profile view of the field. Since we do not have access to ground truth soccer trajectories, we decided to make our own soccer simulator, which we call Soccer World
, using the Unity game engine. The location of each player is determined by a handdesigned “AI”, which is a probabilistic decision tree based on the player’s current state, the states of other players and the ball. The players can take actions such as kicking the ball and tackling. The ball is driven by the players’ actions. Each game lasts 5 minutes. For each game, a player with the same id is assigned to a random position to play, except for the two goal keepers. We render the players using offtheshelf human models from Unity. The identities of the players are colorcoded on their shirts. The camera tracks the ball, and shows a partial view of the field at each frame. See Figure
2(a) for some visualizations. We create 700 videos for training, and 300 videos for test. We apply a random sliding window of 10 seconds to sample the training data. The test videos are uniformly segmented into 10second clips, resulting in 9000 total test examples. We plan to release the videos along with the game engine after publication of the paper. ^{‡}^{‡}‡Video samples can be found at bit.ly/2E3qg6F4.2 Experimental setup
The first task we evaluate is inferring the current state (which we define to be the 2d location) of the all the objects (players and ball). To do this, we replace the discrete predictions with the corresponding realvalued coordinates by using weighted averaging. We then compute normalized distance between the ground truth locations and the predicted locations of all objects at each step.
The second task is predicting future states. Since there are many possible futures, using loss does not make sense, either for training or evaluation (c.f., (Mathieu et al., 2016)). Instead we evaluate the negative loglikelihood on the discretized predictions, which can better capture the multimodal nature of the problem. (This is similar to the perplexity measure used to evaluate language models, except we apply it to each object separately and sum the results.)
For each task, we consider the following models: Visual only: standalone visual encoder without any recurrent neural network as backbone. For future prediction, we follow Felsen et al. (2017) and directly predict future locations from last observed visual features. RNN: standard RNN, the hidden states are globally shared by all agents. VRNN: standard VRNN. IndepRNN: one RNN per agent, each RNN runs independently without any interaction. SocialRNN: one RNN per agent, agents interact via the pooling mechanism in Social GAN (Gupta et al., 2018). GraphRNN: one RNN per agent, with a graph interaction network. GraphVRNN: the full model, which adds stochasticity to GraphRNN. All models have the same architectures for visual encoders and state decoders.
Our visual encoder is based on ResNet18 (He et al., 2016)
, we use the first two blocks of ResNet to maintain spatial resolution, and then aggregate the feature map with max pooling. The encoder is pretrained on visible players, and then finetuned for each baseline. We find this step important to stabilize training. For the soccer data, we downsample the video to 4 FPS, and treat 4 frames (1 second) as one step. We consider 10 steps in total, 6 observed, 4 unobserved. We set the size of GRU hidden states to 128 for all baselines. The state decoder is a 2layer MLP. For basketball data, we set every 5 frames as one step, and consider 10 steps as well. The size of GRU hidden states is set to 128. The parameters of the VRNN backbone and state decoders are shared across all agents, while each agent has its own visual encoder as “identifier”. For all experiments, we use the standard momentum optimizer. The models are trained on 6 V100 GPUs with synchronous training with batch size of 8 per GPU, we train the model for 80K steps on soccer and 40K steps on basketball. We use a linear learning rate warmup schedule for the first 1K steps, followed by a cosine learning rate schedule. The hyper parameters, such as the base learning rate and the KL divergence weight
, are tuned on a holdout validation set.Method  
Visual only  0.192  0.190  0.188  0.188  0.190  0.081  0.078  0.077  0.077  0.080 
RNN  0.189  0.167  0.164  0.164  0.163  0.067  0.061  0.037  0.047  0.037 
VRNN  0.185  0.168  0.167  0.166  0.166  0.074  0.058  0.035  0.044  0.035 
IndepRNN  0.183  0.169  0.160  0.159  0.157  0.071  0.066  0.038  0.046  0.036 
SocialRNN  0.186  0.167  0.160  0.156  0.152  0.069  0.057  0.036  0.043  0.037 
GraphRNN  0.189  0.169  0.159  0.153  0.147  0.068  0.055  0.034  0.041  0.034 
GraphVRNN  0.184  0.165  0.153  0.149  0.143  0.062  0.052  0.025  0.034  0.024 
Method  Ball  Player  
Visible  Hidden  Visible  Hidden  
Visual only  0.011  0.277  0.016  0.315 
RNN  0.008  0.266  0.028  0.266 
VRNN  0.009  0.260  0.029  0.250 
IndepRNN  0.009  0.279  0.039  0.262 
SocialRNN  0.010  0.260  0.044  0.242 
GraphRNN  0.011  0.251  0.034  0.240 
GraphVRNN  0.006  0.227  0.013  0.214 
Method  Perm. Inv.  Inter act  Stoch astic  Soccer  Basket ball 

Prior  No  No  No  
Visual only  No  No  No  
RNN  No  Yes  No  
VRNN  No  Yes  Yes  
IndepRNN  Yes  No  No  
SocialRNN  Yes  Yes  No  
GraphRNN  Yes  Yes  No  
GraphVRNN  Yes  Yes  Yes 
4.3 Quantitative results
Basketball.
The right half of Table 1 shows the average normalized distance between the true and predicted location of all the agents for the basketball data. (Error bars are not shown, but variation across trials is quite small.) We see that the GraphVRNN error generally goes down over time, as the system integrates evidence from multiple frames. (Recall that one of the agents becomes “invisible” every 10 frames or 2 steps, so just looking at the current frame is insufficient.)
The visualonly baseline has more or less constant error, which is expected since it does not utilize information from earlier observations. All other methods outperform the visualonly baseline. We can see that stochasticity helps (VRNN better than RNN, GraphVRNN better than GraphRNN), and GraphRNN is better than vanilla RNN. As most of the players are visible in the basketball videos, we also report performance for visible and hidden (occluded) agents in Table 2. As expected, the distances for visible agents are very low, since the localization task is trivial for this data. When the agents are hidden, GraphVRNN significantly outperforms other baselines. Again, we observe that stochasticity helps. We also find that graph network is a better interaction model than social pooling, both outperform IndepRNN.
Soccer.
The left half of Table 1 shows the average normalized distance between the true and predicted location of all the agents for the soccer data as a function of time. The results are qualitatively similar to the basketball case, although in this setting the distances are higher since the vision problem is much harder. However, the gains from adding stochasticity to the GraphRNN are smaller in this setting. We believe the reason for this is that the dynamics of the agents in soccer world is much more predictable than in the basketball case, because our simulator is not very realistic. (This will become more apparent in Section 4.4.)
Forecasting.
In this section, we assess the ability of the models to predict the future. In particular, the input is the last 6 steps, and we forecast the next 4 steps. Since the future is multimodal, we do not use error, but instead we compute the loglikelihood of the discretized ground truth locations, normalized by the performance of random guessing^{§}^{§}§ In the case of variational models, we report the ELBO, rather than the true log likelihood. Technically speaking, this makes the numbers incomparable across models. However, all the inference networks have the same structure, so we believe the tightness of the lower bound will be comparable. , which is a uniform prior over all locations. The prior baseline trains the decoder based on constant zero input vectors, thus reflecting the prior marginal distribution over locations. The visual only baseline predicts all 4 future steps based on last observed visual feature, as is in Felsen et al. (2017). The results are shown in Table 3. Not surprisingly, we see that modeling the interaction between the agents helps, and adding stochasticity to the latent dynamics also helps.
4.4 Qualitative results
This section shows some qualitative visualizations of the belief state of the system (i.e., the marginal distributions ) over time. (Note that this is different from the beliefs over the internal states, , which are uninterpretable.) We consider the tracking case, where , and the forecasting case, where . We include visualizations of sampled trajectories in the Appendix.
Basketball.
In Figure 4, we visualize the belief state for a single agent in the basketball dataset as a function of time. More precisely, we visualize the output of the decoder, , as a heatmap, where are different stochastic samples of the latent state drawn from the dynamical prior. When we roll forward in time, we sample a specific value , and use this to update each . When we forecast the future, we set . During tracking, when visual evidence is available, we see that the entropy of the belief state reduces over time, and is consistent across different draws. However, when forecasting, we see increased entropy, reflecting uncertainty in the dynamics. Furthermore, the trajectory of the belief state differs along different possible future scenarios (values of ). Figure 6 shows the examples trajectories for five attacking players generated by GraphVRNN and baseline methods. In Figure 5(a), we can see that the samples generated by GraphVRNN are consistent with ground truth for the first 5 to 6 steps (observed), and are diverse for the next 4 steps (not observed). Figure 5(b) shows the trajectories generated for the same example by IndepRNN and vanilla RNN, which are cluttered or shaky.
Soccer.
Figure 5 (left) shows the belief states for the soccer domain for three different kinds of agents: a regular player (top row), the goal keeper (middle row), and the ball (bottom row). For the player, we see that the initial belief state reflects the 442 pattern over possible initial locations of the players. In frame 3, enough visual evidence has been accumulated to localize the player to one of two possible locations. For the future prediction, we draw a single stochastic sample (to break symmetry), and visualize the resulting belief states. We see that the model predicts the player will start moving horizontally (as depicted by the heatmap diffusing), since he is already at the top edge of the field. The goal keeper beliefs are always deterministic, since in the simulated game, the movement of the goalie is below the quantization threshold. The ball is tracked reliably (since the camera is programmed to track the ball), but future forecasts start to diffuse. Figure 5 (right) shows similar results for the basketball domain. We see that the dynamics of the players and ball are much more complex than in our soccer simulator.
5 Conclusions and future work
We have presented a method that learns to integrate temporal information with partially observed visual evidence, based on graphstructured VRNNs, and shown that it outperforms various baselines on two simple datasets. In the future, we would like to consider more challenging datasets, such as real sports videos. We would also like to reduce the dependence on labeled data, perhaps by using some form of selfsupervised learning.
Acknowledgements.
We thank Raymond Yeh and Bo Chang for helpful discussions.
References
 Alahi et al. (2016) Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li FeiFei, and Silvio Savarese. Social LSTM: Human Trajectory Prediction in Crowded Spaces. In CVPR, 2016.
 Alemi et al. (2018) Alexander A Alemi, Ben Poole, Ian Fischer, Joshua V Dillon, Rif A Saurous, and Kevin Murphy. Fixing a broken ELBO. In ICML, 2018.
 Babaeizadeh et al. (2018) Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. In ICLR, 2018.
 BarShalom et al. (2011) Yaakov BarShalom, Peter K Willett, and Xin Tian. Tracking and Data Fusion: A Handbook of Algorithms. Yaakov BarShalom, 2011.
 Battaglia et al. (2016) Peter W Battaglia, Razvan Pascanu, Matthew Lai, Danilo J Rezende, and Koray Kavukcuoglu. Interaction networks for learning about objects, relations and physics. In NeurIPS, 2016.
 Battaglia et al. (2018) Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro SanchezGonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulcehre, Francis Song, Andrew Ballard, Justin Gilmer, George Dahl, Ashish Vaswani, Kelsey Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matt Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. arXiv:1806.01261, 2018.
 Bengio et al. (2015) Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In NeurIPS, 2015.
 Bowman et al. (2016) Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy Bengio. Generating sentences from a continuous space. In CoNLL, 2016.
 Buesing et al. (2018) Lars Buesing, Theophane Weber, Sebastien Racaniere, S. M. Ali Eslami, Danilo Rezende, David P. Reichert, Fabio Viola, Frederic Besse, Karol Gregor, Demis Hassabis, and Daan Wierstra. Learning and querying fast generative models for reinforcement learning. arXiv:1802.03006, 2018.
 Chang et al. (2017) Michael B Chang, Tomer Ullman, Antonio Torralba, and Joshua B Tenenbaum. A compositional objectbased approach to learning physical dynamics. In ICLR, 2017.
 Chung et al. (2015) Junyoung Chung, Kyle Kastner, Laurent Dinh, Kratarth Goel, Aaron C Courville, and Yoshua Bengio. A recurrent latent variable model for sequential data. In NeurIPS, 2015.
 Denton & Fergus (2018) Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018.
 Dequaire et al. (2018) Julie Dequaire, Peter Ondrúška, Dushyant Rao, Dominic Wang, and Ingmar Posner. Deep tracking in the wild: Endtoend tracking using recurrent neural networks. The International Journal of Robotics Research, 2018.
 Ehrhardt et al. (2017) Sebastien Ehrhardt, Aron Monszpart, Niloy Mitra, and Andrea Vedaldi. Taking visual motion prediction to new heightfields. arXiv:1712.09448, 2017.
 Felsen et al. (2017) Panna Felsen, Pulkit Agrawal, and Jitendra Malik. What will happen next? forecasting player moves in sports videos. In ICCV, 2017.
 Fraccaro et al. (2016) Marco Fraccaro, Søren Kaae Sønderby, Ulrich Paquet, and Ole Winther. Sequential neural models with stochastic layers. In NeurIPS, 2016.
 Fragkiadaki et al. (2016) Katerina Fragkiadaki, Pulkit Agrawal, Sergey Levine, and Jitendra Malik. Learning visual predictive models of physics for playing billiards. In ICLR, 2016.
 Goyal et al. (2017) Anirudh Goyal, Alessandro Sordoni, MarcAlexandre Côté, Nan Rosemary Ke, and Yoshua Bengio. Zforcing: Training stochastic recurrent networks. In NeurIPS, 2017.
 Gupta et al. (2018) Agrim Gupta, Justin Johnson, Li FeiFei, Silvio Savarese, and Alexandre Alahi. Social GAN: Socially acceptable trajectories with generative adversarial networks. In CVPR, 2018.
 Haarnoja et al. (2016) Tuomas Haarnoja, Anurag Ajay, Sergey Levine, and Pieter Abbeel. Backprop KF: Learning discriminative deterministic state estimators. In nips, 2016.
 He et al. (2016) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.
 Hoshen (2017) Yedid Hoshen. VAIN: Attentional multiagent predictive modeling. arXiv:1706.06122, 2017.

Karl et al. (2017)
Maximilian Karl, Maximilian Soelch, Justin Bayer, and Patrick van der Smagt.
Deep variational bayes filters: Unsupervised learning of state space models from raw data.
In ICLR, 2017.  Kingma & Welling (2014) Diederik Kingma and Max Welling. Autoencoding variational bayes. In ICLR, 2014.
 Kipf et al. (2018) Thomas Kipf, Ethan Fetaya, KuanChieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In ICML, 2018.

Kitani et al. (2017)
Kris M. Kitani, DeAn Huang, and WeiChiu Ma.
Activity forecasting.
In
Group and Crowd Behavior for Computer Vision
. Elsevier, 2017.  Kosiorek et al. (2018) Adam R Kosiorek, Hyunjik Kim, Ingmar Posner, and Yee Whye Teh. Sequential attend, infer, repeat: Generative modelling of moving objects. arXiv:1806.01794, 2018.
 Krishnan et al. (2017) Rahul G Krishnan, Uri Shalit, and David Sontag. Structured inference networks for nonlinear state space models. In AAAI, 2017.
 Lee et al. (2018) Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv:1804.01523, 2018.
 Lee et al. (2017) Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B Choy, Philip H S Torr, and Manmohan Chandraker. DESIRE: Distant future prediction in dynamic scenes with interacting agents. In CVPR, 2017.
 Lerer et al. (2016) Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example. In ICML, 2016.
 Manafifard et al. (2017) M Manafifard, H Ebadi, and H Abrishami Moghaddam. A survey on player tracking in soccer videos. CVIU, 2017.
 Mathieu et al. (2016) Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multiscale video prediction beyond mean square error. In ICLR, 2016.

Mottaghi et al. (2016)
Roozbeh Mottaghi, Hessam Bagherinezhad, Mohammad Rastegari, and Ali Farhadi.
Newtonian scene understanding: Unfolding the dynamics of objects in static images.
In CVPR, 2016.  Rematas et al. (2018) Konstantinos Rematas, Ira KemelmacherShlizerman, Brian Curless, and Steve Seitz. Soccer on your tabletop. In CVPR, 2018.
 SanchezGonzalez et al. (2018) Alvaro SanchezGonzalez, Nicolas Heess, Jost Tobias Springenberg, Josh Merel, Martin Riedmiller, Raia Hadsell, and Peter Battaglia. Graph networks as learnable physics engines for inference and control. In ICML, 2018.
 Sanguesa (2017) Adria Arbues Sanguesa. Identifying basketball plays from sensor data; towards a LowCost automatic extraction of advanced statistics. Master’s thesis, Aarlborg, 2017.
 Santoro et al. (2017) Adam Santoro, David Raposo, David G Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and Tim Lillicrap. A simple neural network module for relational reasoning. In NeurIPS, 2017.
 Tulyakov et al. (2018) Sergey Tulyakov, MingYu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In CVPR, 2018.

van Steenkiste et al. (2018)
Sjoerd van Steenkiste, Michael Chang, Klaus Greff, and Jürgen Schmidhuber.
Relational neural expectation maximization: Unsupervised discovery of objects and their interactions.
In ICLR, 2018.  Veličković et al. (2018) Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In ICLR, 2018.
 Vondrick et al. (2016a) Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Anticipating visual representations from unlabeled video. In CVPR, 2016a.
 Vondrick et al. (2016b) Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016b.
 Walker et al. (2016) Jacob Walker, Carl Doersch, Abhinav Gupta, and Martial Hebert. An uncertain future: Forecasting from static images using variational autoencoders. In ECCV, 2016.
 Watters et al. (2017) Nicholas Watters, Andrea Tacchetti, Theophane Weber, Razvan Pascanu, Peter Battaglia, and Daniel Zoran. Visual interaction networks. In NeurIPS, 2017.
 Wu et al. (2017) Jiajun Wu, Erika Lu, Pushmeet Kohli, Bill Freeman, and Josh Tenenbaum. Learning to see physics via visual deanimation. In NeurIPS, 2017.
 Xie et al. (2018) Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speedaccuracy tradeoffs in video classification. In ECCV, 2018.
 Xue et al. (2016) Tianfan Xue, Jiajun Wu, Katherine Bouman, and Bill Freeman. Visual dynamics: Probabilistic future frame synthesis via cross convolutional networks. In NeurIPS, 2016.
 Zhan et al. (2018) Eric Zhan, Stephan Zheng, Yisong Yue, Long Sha, and Patrick Lucey. Generative multiagent behavioral cloning. arXiv:1803.07612, 2018.
Appendix A.1 Sampled soccer trajectories
Figure A1 shows the sampled trajectories for Soccer World. Unlike basketball, Soccer World has a moving camera with limited field of view. We observe that the trajectories for the first few observed steps are quite shaky since only a few players have been observed. We thus show the trajectories from . From Figure A1, we can see that the trajectories generated by RNN are very shaky. GraphVRNN generates much better trajectories, but some of the players are assigned incorrect identities, or incorrect locations. We conjecture that this issue can be mitigated by providing longer visual inputs to the model, such that most of the players could be observed at some point of the videos.
Comments
There are no comments yet.