Context-aware Human Motion Prediction

TL;DR

This paper introduces a novel context-aware human motion prediction model that incorporates interactions with objects and other humans using a semantic graph and graph attention, outperforming traditional RNN-based methods.

Contribution

The paper presents a new architecture combining semantic graphs and RNNs for context-aware human motion prediction, explicitly modeling interactions with objects and humans.

Findings

Outperforms baseline models without context information.

Effectively models human-object and human-human interactions.

Improves prediction accuracy in complex scenarios.

Abstract

The problem of predicting human motion given a sequence of past observations is at the core of many applications in robotics and computer vision. Current state-of-the-art formulate this problem as a sequence-to-sequence task, in which a historical of 3D skeletons feeds a Recurrent Neural Network (RNN) that predicts future movements, typically in the order of 1 to 2 seconds. However, one aspect that has been obviated so far, is the fact that human motion is inherently driven by interactions with objects and/or other humans in the environment. In this paper, we explore this scenario using a novel context-aware motion prediction architecture. We use a semantic-graph model where the nodes parameterize the human and objects in the scene and the edges their mutual interactions. These interactions are iteratively learned through a graph attention layer, fed with the past observations, which now include both object and human body motions. Once this semantic graph is learned, we inject it to a standard RNN to predict future movements of the human/s and object/s. We consider two variants of our architecture, either freezing the contextual interactions in the future of updating them. A thorough evaluation in the "Whole-Body Human Motion Database" shows that in both cases, our context-aware networks clearly outperform baselines in which the context information is not considered.

Tables

Table 1. Table 1 : Class-specific models results. In this table, every action is independently trained. The results report the mean Euclidean error (in mm), for the 2s prediction of the human motion (top) and object motion (bottom). In all cases, 1s of past observations is provided. The context-based models we propose in this paper are those with the suffixes “OMP” and “LI”. They provide the best results in most sequences.

Human motion	Passing objects				Grasping objects				Cutting food				Mixing objects				Cooking
Time (s)	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2
ZV[43]	34	81	120	153	89	222	333	421	54	132	198	258	102	262	396	495	24	53	70	80
RNN[43]	50	99	132	162	82	158	211	254	48	103	140	180	68	135	190	226	27	54	65	71
QuaterNet	62	145	211	267	208	209	248	292	87	211	308	389	192	237	296	345	39	87	121	144
C-RNN	47	102	141	177	76	149	203	247	49	100	124	158	70	158	214	247	26	53	63	69
C-RNN+OMP	53	99	127	155	128	154	197	239	49	96	121	149	61	127	168	199	29	55	65	70
C-RNN+LI	43	89	117	142	72	141	188	230	47	92	117	147	72	145	194	219	27	53	63	69
C-RNN+OMP+LI	44	89	116	142	115	156	204	251	48	95	121	147	77	152	195	219	26	53	63	68
Object motion	Passing objects				Grasping objects				Cutting food				Mixing objects				Cooking
Time (s)	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2
ZV	48	118	181	237	65	152	226	289	29	70	104	132	50	126	188	229	16	33	44	53
RNN	49	107	154	198	64	139	201	257	29	70	105	134	47	113	166	199	17	36	48	58
C-RNN+OMP	44	92	122	150	55	103	136	167	31	64	83	97	29	65	90	110	15	33	46	56
C-RNN+OMP+LI	44	91	119	142	58	112	152	186	29	62	81	92	51	106	145	171	16	34	46	55

Table 2. Table 2 : Training with all actions simultaneously. For each method we train a single model using all actions simultaneously. See also caption in Table  1 .

All	Human motion prediction
Time (s)	0.1	0.2	0.3	0.4	0.5	0.6	0.7	0.8	0.9	1	1.1	1.2	1.3	1.4	1.5	1.6	1.7	1.8	1.9	2
ZV [43]	24	46	67	87	106	125	143	160	176	190	205	219	231	244	256	267	279	290	300	310
RNN [43]	29	44	57	68	78	87	96	104	113	121	128	136	143	150	157	164	171	177	184	191
C-RNN	27	46	58	69	79	87	96	104	113	121	129	137	144	152	160	166	174	181	188	196
C-RNN+OMP	46	83	76	82	87	95	101	108	116	123	131	138	146	153	160	167	174	182	189	197
C-RNN+LI	21	39	52	63	72	80	89	97	104	111	118	125	131	137	144	150	157	163	170	177
C-RNN+OMP+LI	39	77	77	76	80	87	94	101	108	114	120	126	133	139	145	151	158	165	171	178
All	Object motion prediction
ZV	13	25	35	44	52	60	68	77	84	90	96	102	109	115	120	125	131	135	140	144
RNN	15	28	38	46	53	60	68	74	80	85	91	97	102	107	112	117	121	125	130	135
C-RNN+OMP	15	26	36	44	50	55	61	67	73	79	84	89	94	99	104	108	113	117	121	125
C-RNN+OMP+LI	16	29	39	46	52	57	63	69	75	79	84	88	93	97	101	105	110	114	117	121

	Passing objects	Grasping/ leaving object	Cutting food	Mixing objects	Cooking	All
# objects	4	5	6	9	12	15
# people	2	1	1	1	1	1/2
# videos	18	36	10	17	35	190
# frames	30k	31k	11k	14k	54k	198k

Table 4. Table 3 : Robustness to noise in Human and Object Motion Prediction. Average performance of the principal models when using the original test set (Noise-free input), compared to their performance when seeing noisy observations.

	Noise-free input				Noisy input (HMP)
Model	HMP Models				25 mm				50 mm				100 mm
Time (s)	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2	0.5	1	1.5	2
ZV [43]	61	150	223	281	70	156	227	284	87	168	236	292	122	195	260	313
RNN [43]	55	110	148	179	72	122	156	185	83	127	162	192	112	184	214	226
C-RNN+LI	52	104	136	161	58	107	139	166	69	113	147	175	99	136	175	208
C-RNN+OMP+LI	56	109	140	165	62	111	142	167	71	116	146	171	103	132	162	187
Model	OMP Models				25 mm				50 mm				100 mm
ZV	42	100	149	188	53	106	154	191	66	118	164	199	106	151	187	223
RNN	41	93	135	169	52	99	141	174	69	113	151	183	105	142	181	208
C-RNN+OMP+LI	40	81	109	129	51	88	115	134	67	100	126	144	106	132	156	172

Table 5. Table 4 : Mean and Std of prediction errors (mm) on the CMU dataset. Our Context-aware model C-RNN+LI outperforms baselines even though context only consists of two people.

CMU MoCap Dataset
Time (s)	0.5		1.0		1.5		2.0
	$\mu$	$\sigma$	$\mu$	$\sigma$	$\mu$	$\sigma$	$\mu$	$\sigma$
ZV	127	32	271	66	374	86	460	97
RNN	125	28	267	58	378	77	477	92
QuaterNet	138	26	279	58	378	82	466	95
C-RNN+LI	124	27	257	53	352	65	435	78

Equations

$$\mathcal{M}:\{P_{t_o:t-1},O_{t_o:t-1},T\}\to \{P_{t:t_f},O_{t:t_f},I_{t_o:t_f}\},$$

$$R_i=\sigma (\sum _j^N{\tilde{A}}_{ij}W[x_i;x_i-x_j]).$$

$$H_i^{t+1}=\text{RNN}(R_i^t,H_i^t).$$

$$A_{ij}^t=g(H_i^t,H_i^t-H_j^t),$$

Full text

In this paper, we explore this scenario using a novel context-aware motion prediction architecture. We use a semantic-graph model where the nodes parameterize the human and objects in the scene and the edges their mutual interactions. These interactions are iteratively learned through a graph attention layer, fed with the past observations, which now include both object and human body motions. Once this semantic graph is learned, we inject it to a standard RNN to predict future movements of the human/s and object/s. We consider two variants of our architecture, either freezing the contextual interactions in the future of updating them. A thorough evaluation in the “Whole-Body Human Motion Database” [29] shows that in both cases, our context-aware networks clearly outperform baselines in which the context information is not considered.

1 Introduction

The ability to predict and anticipate future human motion based on past observations is essential for interacting with other people and the world around us. While this seems a trivial task for a person, it involves multiple sensory modalities and complex semantic understanding of the environment and the relations between all objects in it. Modeling and transferring this kind of knowledge to autonomous agents would have a major impact in many different fields, mainly in human-robot interaction [30] and autonomous driving [47], but also in motion generation for computer graphics animation [31] or image understanding [10].

The explosion of deep learning, combined with large-scale datasets of human motion such as Human3.6M [24] or the CMU motion capture dataset [34], has led to a significant amount of recent literature that tackles the problem of forecasting 3D human motion from past observations [14, 25, 43, 20, 3, 37, 15, 42, 26, 66]. These algorithms typically formulate the problem as sequence-to-sequence task, in which past observations represented 3D skeleton data are injected to a Recurrent Neural Network (RNN) which then predicts movements in the near future (less than 2 seconds).

Nevertheless, while promising results have been achieved, we argue that the standard definition of the problem used so far lacks an important factor, which is the influence of the rest of the environment on the movement of the person. For instance, if a person is carrying a box, the configuration of the body arms and legs will be highly constrained by the 3D position of that box. Discovering such interrelations between the person and the object/s of the context (or another person he/she is interacting with), and how these interrelations constrain the body motion, is the principal motivation of this paper.

In order to explore this new paradigm, we devise a context-aware motion prediction architecture, that models the interactions between all objects of the scene and the human using a directed semantic graph. The nodes of this graph represent the state of the person and objects (e.g. positional features) and the edges their mutual interactions. These interactions are iteratively learned with the past observations of the human and objects motion and fed into a standard RNN which is then responsible for predicting the future movement of all elements in the scene (for both rigid objects and non-rigid human skeletons). Additionally, we propose a variant of this model that also predicts the evolution of the adjacency matrix representing the interaction between the elements of the scene.

Presumably, one of the reasons why current state-of-the-art has not considered an scenario like ours is because all methods are trained and evaluated on benchmarks (mostly the aforementioned Human3.6M dataset [24]) annotated only with human motion. In this paper, we thoroughly evaluate our approach in the “Whole-Body Human Motion Database” [29], that contains about 200 videos of people performing several tasks and interacting with objects. This dataset is annotated with MoCap data for the humans and rigid displacement for the rest of objects, being thus, a perfect benchmark to validate our ideas. We also evaluate our method in the CMU MoCap database [34] with only two people being tracked. The results obtained in both datasets show that our methodology is able to accurately predict the future motion of people and objects while simultaneously learning very coherent interaction relations. Additionally, all context-aware versions, clearly outperform the baselines which uniquely rely on human past observations of the human (see Fig. 1). Since all previous works evaluate their methods using past observations of ground truth skeletons, we finally discuss the applicability of state-of-art motion prediction methods, with an ablation study of our models and baselines when considering noisy observations.

2 Related work

Human motion prediction. Since the release of large-scale MoCap datasets [52, 24, 29], there has been a growing interest in the problem of estimating 3D human pose from single images [5, 52, 49, 60, 56, 55, 44, 57]. More recently, the community is focusing in predicting 3D human motion from past observations. Most approaches build upon RNNs [14, 43, 20, 3, 50, 1] that encode historical motion of the human and predict the future configuration that minimizes different sort of losses. Martinez et al. [43], for instance, minimize the L2 distance and provide one of the baselines in our work. This work also compares against a zero-velocity baseline, which despite steadily predicting the last observed frame, yields very reasonable results under the L2 metric. This phenomenon has been recently discussed by Ruiz et al. [54], that argue that L2 distance is not an appropriate metric to capture the actual distribution of human motion, and that a network trained using only this metric is prone to converge to a mean body pose. To better capture real distributions of human movement, recent approaches use adversarial networks [17, 2] in combination with geometric losses [3, 20, 54, 33].

There exist alternative approaches other than RNNs. For instance, Jain et al. [25] consider a hand-crafted spatial-temporal graph adapted to the skeleton shape. Li et al. [37] use Convolutional Neural Networks to encode and decode skeleton sequences instead of RNNs.

All methods described in this section formulate the human prediction problem without considering the context information. In this paper, we aim to fill this gap.

Rigid 3D object motion prediction. While there is a vast amount of works on 3D object reconstruction [51, 19, 41], detection [9, 18, 11] and tracking [8, 4], only very few approaches address the problem of predicting future rigid motion [6, 63, 32, 59]. Among these, it is worth to mention Byravan et al. [6], that predict the future 3D pose given an image of an object and the action being applied to it. In our case, the action applied to each object is implicitly encoded in the previous observations.

Human-Object Interaction (HOI). Even though our work does not aim to identify Human-Object relationships, we have been inspired by a few papers on this topic. The standard formulation of the problem consists in representing an image with several detected objects and people as a graph encoding the context [22, 45, 53, 39, 16], or some other structured representation [36, 65, 12]. The most recent approaches [37, 53, 22] extract features of the detected entities using some image-based classification CNNs. Then, they compare pairs of features to predict their mutual interaction. Qi et al. [53] refine the representations and predicted interactions in a recursive manner. In this work, we use a similar idea to progressively refine the estimation of the interactions between objects.

Graph-based context reasoning. A few works leverage context information to boost the performance of different tasks[48, 38, 23, 35, 46]. Graph Convolutional Networks (GCNs) [28] were recently proposed for improved semi-supervised classification. Jain et al. [25] used Structural RNNs to model spatio-temporal graphs. Wang et al. [61] propose to use GCNs, in which the interactions between objects depend on the intersection over union of their detected bounding boxes. Chen et al. [10] introduce an approach for image segmentation in which features from a 2D image coordinate space are represented in a graph reasoning space.

3 Problem formulation

Recent methods for human motion prediction consist of a model $\mathcal{M}$, typically a deep neural network, that encodes motion from time $t_{o}$ until $t-1$. The goal is then to predict future human motion until $t_{f}$, namely $P_{t:t_{f}}$, where $P$ stands for the human pose represented by 3D joint coordinates. Previous approaches have formulated the problem as $\mathcal{M}:P_{t_{o}:t-1}\rightarrow P_{t:t_{f}}$, i.e. future motion is estimated only from past observations. In this paper, we conjecture that future motion is also driven by the context and the action the human is performing. We therefore consider other objects $O$ of type $T$ in the scene with which the human may interact. The objects can be other people or any object in the scene. We will design our approach to be able to predict the motion of such objects of the context.

Additionally, the influence that objects will have in the future motion of other objects is unclear. Thus, we also aim to build a model that learns these interactions in an unsupervised manner. Considering all this, we reformulate our problem as the estimation of the following mapping:

[TABLE]

where $I$ corresponds to the predicted interactions.

4 Approach

Figure 2 shows the main architecture used in this work. It consists of two branches that separately process human motion and object relationships. We use the latter to obtain a representation for all the observed entities, including the human, which we then use to predict both human and object motion prediction. We next describe these two branches.

4.1 Human motion branch

This branch builds upon the RNN network proposed by Martinez et al. [43]. This model, depicted in blue in Figure 2, is based on a residual architecture [21] that, at each step, uses a fully connected layer to predict the velocity of the body joints. As in a typical sequence-to-sequence network, the predictions are fed to the next step.

4.2 Context branch

The context information is represented using a directed graph structure where each node denotes an object or person. We then store a state for each entity and frame, encoding context information relevant to each node. These states are iteratively refined as new observations are processed.

Object representation. At each frame $t$, we define a matrix $X_{t}\ \in\mathbb{R}^{N\times F_{0}}=\left[O_{t},T_{t},P_{t}\right]$ that gathers the representation of all $N$ nodes. $F_{0}$ is the length of the state vector of each node. This state vector contains the object 3D bounding box $O_{t}$, their object type $T$ as a one-hot vector, and the joints of the person $P_{t}$. If the node does not correspond to a person, the joints in the representation are set to a zero vector of same size. The object type helps to identify the task the human is performing and the motion defined for that task. By doing this, we aim to capture the semantic difference between the motion of a person when handling a knife or when using a whisk.

Modelling contextual object representations. Recent works on Graph Convolutional Networks (GCNs) [28] have shown very promising results in a variety of problems requiring the manipulation of graph-structured data. In GCNs, a feature vector of a certain node $R_{i}$ is expressed as a function of other nodes $x$, as $R_{i}=\sigma(\sum^{N}_{j}\tilde{A}_{ij}Wx_{j})$, where $W$ are trainable weights, $\sigma$ is an activation and $N$ the number of nodes of the graph connected to the $i$-th node. $\tilde{A}\in\mathbb{R}^{N\times N}$ is a normalized weighted adjacency matrix that defines interactions between nodes.

Graph Attention Networks (GATs) [58] have been proposed as an extension of GCNs, and introduce an attention model on every graph node. In this paper we also investigate the use of Edge Convolutions [62], which are indeed very similar to GATs. In ECs the update rule for a feature vector of each entity considers the representations of other relevant objects as follows:

[TABLE]

The intuition behind this equation is that $x_{i}$ encodes a global representation of the node, while $x_{i}-x_{j}$ provides local information. EC proposes combining both types of information in an asymmetric graph function.

We keep track of the context representations during all observations through a second RNN. Each node on the scene has a hidden state $H$ that is updated every frame $t$:

[TABLE]

Learning interactions. As we shall see in the experimental section, we initially evaluate a simplified version of our Context-RNN (C-RNN) that uses a heuristic to define the adjacency matrices, setting $A_{ij}=1$ if the center of gravity of objects $i$ and $j$ is closer than 1 meter.

In practice, interactions between entities are not known a priory, and furthermore, they change over time. Our goal is to automatically learn these changing interactions with no supervision. For this purpose we devise an iterative process in which, for the first frame, we set $A$ to a diagonal matrix, i.e. $\tilde{A}_{t_{0}}=I_{N}$, meaning that the initial hidden representation of every object depends only on itself. We then predict the value of the interaction between two objects given the hidden state of both. We consider asymmetric weighted adjacency matrices, that for a frame $t$ are estimated as:

[TABLE]

with similar structure as in Eq. 2. The function $g$ represents the output of a neural network layer, in our case a fully connected. We normalize the interactions for each node using a Softmax function, which we shall denote $\tilde{A}$.

Intuitively, we can consider this as a complete graph, where a graph attention mechanism [58] decides on the strength of interactions based on past observations. Note that while existing works typically use binary adjacency matrices from ground truth relationships [28], spatial assumptions [61] or K-NN on node representations [62], in this work we consider a differentiable continuous space of interactions, learned using back-propagation. In the rest of the paper we will denote the models that learn interactions with the suffix “-LI” (e.g. C-RNN+LI).

Object motion prediction. We propose two methods that exploit context at different levels. First, in the blue+brown modules of Fig. 2, we consider a model that reasons about the past context observations and iteratively improves hidden representations. The refined context representation of the human node is concatenated to the baseline branch (in blue) representation at every time step, and used by a fully connected layer to predict human velocity in that step. This is followed by a residual layer that yields skeleton poses.

Our second approach consists of the complete model depicted in Fig. 2 which, apart from past context, predicts object motion for all objects using a residual fully connected layer on each object hidden state. Analogous to the human motion branch, the predicted positions are forwarded to the next step, allowing to extend the context analysis into the future. The joints in the feature representations for those nodes describing people are also updated with the joint predictions of the human branch.

Additionally, when tracking several people, the human motion branch is repeated for each of them, and the model provides complete future motion for all available entities. In the rest of the document, we will denote the models that predict object motion with the suffix “-OPM”.

5 Implementation details

Our model builds on the residual architecture of Martinez et al. [43] to allow an unbiased comparison with their work. The size of the human and object RNN hidden representations are 1024 and 256, respectively.

After the motion seed, we sample an observation every 100 ms. In all experiments, we encode and decode 10 (1 sec.) and 20 frames (2 sec.) respectively. Larger encoding times did not help in improving the results and significantly increased training time. We augment the train set through random rotation over the height $Z$ in the range $(-180,180]^{\circ}$ and random translation $X,Y\in(-1500,1500)$mm.

We use a similar approach as in [53] to obtain the adjacency matrix. We build a 4D matrix $A$ such that $A_{ij}$ contains the hidden representations $[H_{i}^{t};H_{i}^{t}-H_{j}]$ of nodes $i$ and $j$, extending over the channel dimension. The function $g(\cdot)$ is formed by two Convolutional Layers of output kernel size $1$ to make computation faster. We do not use bias term in these Convolutional layers nor in the Edge Convolutions. Object representations are formed first by the bounding box position, defined by the minimum and maximum 3D Cartesian points.

We train the model to minimize $L2$ distance between the predicted and the actual future motion $\mathcal{L}=||M(P_{t_{o}:t-1})-P_{t:t_{f}}||_{2}$. The model is trained until convergence, using Adam [27] with learning rate of 0.0005, beta1 0.5, beta2 0.99 and batch size 16.

6 Experiments

6.1 Preliminaries

Datasets. Large-scale MoCap datasets [29, 24, 34] provide annotations on the human poses but do not give any annotation about objects of the scene or any relevant context information. Therefore, most recent works on human motion prediction are evaluated without considering context information. Martinez et al. [43] show that for certain cases, even a simple zero-velocity baseline may yield better results than context-less learning models.

To demonstrate the merits of our approach, we leverage on the Whole-Body Human Motion (WBHM) Database [40], a large-scale publicly available dataset containing 3D raw data of multiple individuals and objects. In particular, we use all the activities where human joints are provided and include at least a table. This results in 190 videos and 198K frames, and a total of 15 tracked object classes. We use the raw recordings Vicon files at 100 Hz to obtain the bounding box of each object in each frame, and select 18 joints to represent the human skeleton.

We extract different actions representing different levels of complexity on the contextual information. The statistics of this dataset are the following:

We will report results on both action-specific models and also on models trained with the entire dataset.

We also run experiments on the CMU Mocap Database [34]. We select the actions that include two people interacting, which include 34 videos with different activities like dancing, talking with hand gestures or boxing. In this case, the objects are not annotated, but we will show that context information from the two users is useful to improve over context-less models.

Baselines. We compare our models to the context-less models proposed in [43]. First, we consider the basic residual RNN. We also consider a Zero-Velocity (ZV) baseline that constantly predicts the last observed frame. We also compare to QuaterNet [50] using their available code, to predict absolute motion prediction. For object motion prediction, we also use a ZV and RNN models [43], where the position of an object is defined by its 3D bounding box.

Our models. We run our context-aware models (C-RNN), incrementally adding the main ideas described in the paper.

The basic C-RNN in our experiments uses the spatial heuristic described in Section 4.2 where interactions depend only on the distance between objects. This model processes context during the past frames, and then uses the last hidden state of the human node for human motion prediction at each step. This is extended by additionally predicting object motion (OMP) and recomputing object interaction from the previous assumption on the predicted positions. We then evaluate the efficiency of our model for learning interactions (LI). Like in the previously defined experiments, we evaluate a model that considers past contextual information and a model that prolongs object analysis into the future.

Evaluation metric. Previous works on human motion prediction focus mainly on predicting relative motion [43, 20, 50], using joint angles. However, our model reasons about the full scene and is able to predict absolute motion in Cartesian coordinates. Therefore, we use the mean Euclidean Distance (in mm) between predictions and real future motion, obtained from the unnormalized predictions in the 3D space. For human motion prediction, we take into account the 18 joints defined in the human skeleton. For objects, we consider the eight 3D vertices of their bounding boxes.

6.2 Results on the WBHM Dataset

Quantitative results. Table 1 summarizes the performance of class-specific models trained on different activities. Table 2 provides results at much higher temporal resolution for models trained using all the dataset, reporting the mean Euclidean distance between predictions and ground truth every 100 ms. In all cases, 1 second of past observations is provided and 2 seconds are predicted.

The performance of models that consider a threshold-based binary interaction vary significantly between classes, suggesting they are effectively unable to understand the context as done by models that learn the actual interactions (LI). Notice that even the basic C-RNN does not yet provide a consistent improvement compared to state-of-art models. The same model that additionally learns interactions (C-RNN+LI) obtains a significant boost in most cases. Nonetheless, activities such as passing objects or grasping require attending to items that are at variable distances.

Regarding the complexity of the scene, most improvement comes from scenes with a small number of objects where interactions are well defined and actions are more predictable. For cooking activities, there are several objects in a table next to the human. Different motion options are possible and, as uncertainty grows, the model seems unable to confidently understand interactions. Because of this, context-aware models do not provide such a significant improvement as in previous activities. Considering all actions simultaneously seems to favor even more the context-aware approaches and, specially, those that learn interactions (C-RNN+LI and C-RNN+OPM+LI).

Qualitative results. Figure 3-left shows the motion generation results of our two main models, compared to the baseline [43] on different classes. We did not include the Zero-Velocity baseline as it does not provide interesting motion even though it has remained a difficult baseline on uncertain activities. We have marked some specific frames in which context-aware approaches improve the RNN baseline.

For human motion prediction, poses generated are frequently more semantically-related to their closest objects than context-less models. For instance, as shown in the last action of Figure 3, people holding objects tend to move the relevant hand. For object motion prediction, context-less model predictions hardly move from their original position.

Regarding the interactions predicted by the model, we notice coherent patterns in many activities. For example, drinking videos generate strong Cup-Human relationships. In Figure 4, we represent the average predicted interactions for different actions. These are gathered from the C-RNN+OMP+LI. This model provides more intense Object-Object interactions than C-RNN+LI, which does not need to obtain such meaningful representations for objects as only human contextual representations are used. Note that the models learn to predict interactions that provide information relevant to future pose, and thus improve motion predictions. Interactions here do not necessarily respond to actual action relationships.

We finally study the effect of the Graph architecture in the learned interactions. Graph Attention Networks (GATs) and Edge Convolutions (EC) provide an attention mechanism to measure the interaction strength. Nevertheless, we found that GAT-based networks consider all interactions of similar importance, while EC-based architectures are able to predict a continuous and wide range of attention values. We show this in Figure 5.

6.3 Results on the CMU MoCap Dataset

We train the models again on the CMU MoCap Database, obtaining the results depicted in Table 4. In this setup, the users perform very energetic activities like dancing or boxing, which implies that absolute motion is larger, and error on the CMU MoCap database being in average more than twice that in the former database. In this case, only two nodes are observed in each video for the two people being tracked. Since no information about actions or objects is given, we do not provide results on OMP. However, we find our proposed model C-RNN+LI outperforms all other baselines significantly, specially in the long-term.

6.4 Robustness to noise

All previous works on human motion prediction use ground truth MoCap data as past observations. Nevertheless, real applications will receive joint observations from e.g. human pose estimation models, such as OpenPose [7] or AlphaPose [13, 64], which are prone to suffer from noise and mis-detections, specially under strong occlusions. In these subsection, we therefore evaluate the resilience of our proposed models and previous baselines to noise in the input observations. Predictions are evaluated on the original ground truth data. The 3D coordinates of past observations (both in human and objects positions) are corrupted by additive Gaussian noise $\mathcal{N}(0,\,\sigma^{2})$. In Table  3 we show the results of this experiment, with different values of $\sigma$. Interestingly, the error in the predictions gracefully increases with the noise, but still, our approach performs consistently better than those approaches that do not consider the context information. Indeed, the best context-aware models (C-RNN+LI and C-RNN+OMP+LI) with noise up to $\sigma=50mm$, perform better than context-less baselines with no noise in the input.

7 Conclusion

In this work, we explore a context-aware motion prediction architecture, using a semantic-graph representation where objects and humans are represented by nodes independently of the number of objects or complexity of the environment. We extensively analyze their contribution for human motion prediction. The results observed in different actions suggest that the models proposed are able to understand human activities significantly better than state-of-art models which do not use context, improving both human and object motion prediction.

Acknowledgements: This work has been partially funded by Spanish government with the projects HuMoUR TIN2017-90086-R and the ERA-Net Chistera project IPALM PCI2019-103386. We also thank Nvidia for hardware donation.