Section: New Results
Animation, Autonomous Characters and Digital Storytelling
MimeTIC main research path consists in associating motion analysis and synthesis to enhance the naturalness in computer animation, with applications in movie previsualisation, and autonomous virtual character control. Thus, we pushed example-based techniques in order to reach a good tradeoff between simulation efficiency and naturalness of the results. In 2019, to achieve this goal, MimeTIC continued to explore the use of perceptual studies and model-based approaches, but also began to investigate deep learning, for example to control cameras in Movie previsualization.
VR as a Content Creation Tool for Movie Previsualisation
Participants : Marc Christie [contact] , Quentin Galvane.
This work proposes a VR authoring system which provides intuitive ways of crafting visual sequences in 3D environments, both for expert animators and expert creatives. It is designed in mind to be applied animation and film industries, but can find broader applications (eg. in multimedia content creation). Creatives in animation and film productions have forever been exploring the use of new means to prototype their visual sequences before realizing them, by relying on hand-drawn storyboards, physical mockups or more recently 3D modelling and animation tools. However these 3D tools are designed in mind for dedicated animators rather than creatives such as film directors or directors of photography and remain complex to control and master. The proposed system is designed to reflect the traditional process through (i) a storyboarding mode that enables rapid creation of annotated still images, (ii) a previsualisation mode that enables the animation of the characters, objects and cameras, and (iii) a technical mode that enables the placement and animation of complex camera rigs (such as cameras cranes) and light rigs. Our methodology strongly relies on the benefits of VR manipulations to re-think how content creation can be performed in this specific context, typically how to animate contents in space and time. As a result, the proposed system is complimentary to existing tools, and provides a seamless back-and-forth process between all stages of previsualisation. We evaluated the tool with professional users to gather experts’ perspectives on the specific benefits of VR in 3D content creation .
Deep Learning Techniques for Camera Trajectories
Participant : Marc Christie [contact] .
Designing a camera motion controller which places and moves virtual cameras in relation with contents in a cinematographic way is a complex and challenging task. Many cinematographic rules exist, yet practice shows there are significant stylistic variations in how these can be applied. While contributions have attempted to encode rules by hand, this work is the very first to propose an end-to-end framework that automatically learns from real and synthetic movie sequences how the camera behaves in relation with contents. Our deep-learning framework extracts cinematic features of movies through a novel feature estimator trained on synthetic data, and learns camera behaviors from those extracted features, through the design of a Recurrent Neural Network (RNN) with a Mixture of Experts (MoE) gating mechanism. This cascaded network is designed to capture important variations in camera behaviors while ensuring the generalization capacity in the learning of similar behaviors. We demonstrate the features of our framework through experiments that highlight (i) the quality of our cinematic feature extractor (ii) the capacity to learn ranges of behaviors through the gating mechanism, and (iii) the ability to analyse the camera behaviors from a given input sequence, and automatically re-apply these behaviors on new virtual contents, offering exciting new possibilities towards a deeper understanding of cinematographic style and enhanced possibilities in transferring style from real to virtual. The work is a collaboration with the Beijing Film Academy in China.
Efficient Visibility Computation for Camera Control
Participants : Marc Christie [contact] , Ludovic Burg.
Efficient visibility computation is a prominent requirement when designing automated camera control techniques for dynamic 3D environments; computer games, interactive storytelling or 3D media applications all need to track 3D entities while ensuring their visibility and delivering a smooth cinematographic experience. Addressing this problem requires to sample a very large set of potential camera positions and estimate visibility for each of them, which in practice is intractable. In this work, we introduce a novel technique to perform efficient visibility computation and anticipate occlusions. We first propose a GPU-rendering technique to sample visibility in Toric Space coordinates – a parametric space designed for camera control. We then rely on this visibility evaluation to compute an anticipation map which predicts the future visibility of a large set of cameras over a specified number of frames. We finally design a camera motion strategy that exploits this anticipation map to maximize the visibility of entities over time. The key features of our approach are demonstrated through comparison with classical ray-casting techniques on benchmark environments, and through an integration in multiple game-like 3D environment with heavy sparse and dense occluders.
Analysing and Predicting Inter-Observer Gaze Congruency
Participant : Marc Christie [contact] .
In trying to better understand film media, we have been recently exploring the relation between the distribution of gaze states and the features of images, with the objective of establishing correlations to understand how films manipulate users gaze (and how gaze can be manipulated be re-editing film sequences). According to the literature regarding visual saliency, observers may exhibit considerable variations in their gaze behaviors. These variations are influenced by aspects such as cultural background, age or prior experiences, but also by features in the observed images. The dispersion between the gaze of different observers looking at the same image is commonly referred as inter-observer congruency (IOC). Predicting this congruence can be of great interest when it comes to study the visual perception of an image. We introduce a new method based on deep learning techniques to predict the IOC of an image . This is achieved by first extracting features from an image through a deep convolutional network. We then show that using such features to train a model with a shallow network regression technique significantly improves the precision of the prediction over existing approaches.
Deep Saliency Models: the Quest for the Loss Function
Participant : Marc Christie [contact] .
Following our idea of understanding gaze patterns in movie watching, and predicting these gaze patterns on sequences, we have been exploring the influence of loss functions in learning the visual saliency. Indeed, numerous models in the literature present new ways to design neural networks, to arrange gaze pattern data, or to extract as much high and low-level image features as possible in order to create the best saliency representation. However, one key part of a typical deep learning model is often neglected: the choice of the loss function. In this work, we explore some of the most popular loss functions that are used in deep saliency models . We demonstrate that on a fixed network architecture, modifying the loss function can significantly improve (or depreciate) the results, hence emphasizing the importance of the choice of the loss function when designing a model. We also introduce new loss functions that have never been used for saliency prediction to our knowledge. And finally, we show that a linear combination of several well-chosen loss functions leads to significant improvements in performances on different datasets as well as on a different network architecture, hence demonstrating the robustness of a combined metric.
Contact Preserving Shape Transfer For Rigging-Free Motion Retargeting
Participants : Franck Multon [contact] , Jean Basset.
In 2018, we introduced the idea of context graph to capture the relationship between body parts surfaces and enhance the quality of the motion retargetting problem. Hence, it becomes possible to retarget the motion of a source character to a target one while preserving the topological relationship between body parts surfaces. However this approach implies to strictly satisfy distance constraints between body parts, whereas some of them could be relaxed to preserve naturalness. In 2019, we introduced a new paradigm based on transfering the shape instead of encoding the pose constraints to tackle this problem .
Hence, retargeting a motion from a source to a target character is an important problem in computer animation, as it allows to reuse existing rigged databases or transfer motion capture to virtual characters. Surface based pose transfer is a promising approach to avoid the trial-and-error process when controlling the joint angles. The main contribution of this work is to investigate whether shape transfer instead of pose transfer would better preserve the original contextual meaning of the source pose. To this end, we propose an optimization-based method to deform the source shape+pose using three main energy functions: similarity to the target shape, body part volume preservation, and collision management (preserve existing contacts and prevent penetrations). The results show that our method is able to retarget complex poses, including several contacts, to very different morphologies. In particular, we introduce new contacts that are linked to the change in morphology, and which would be difficult to obtain with previous works based on pose transfer that aim at distance preservation between body parts. These preliminary results are encouraging and open several perspectives, such as decreasing computation time, and better understanding how to model pose and shape constraints.
The Influence of Step Length to Step Frequency Ratio on the Perception of Virtual Walking Motions
Participants : Ludovic Hoyet [contact] , Benjamin Niay, Anne-Hélène Olivier.
Synthesizing walking motions that look realistic and diverse is a challenging task in animation, and even more when the target is to create realistic motions for large group of characters. Indeed, in order to keep a good trade-off between computational costs and realism, biomechanical constraints of human walk are not always fulfilled. In pilot experiments , , we have therefore started to investigate the ability of viewers to identify an invariant parameter of human walking named the walk ratio, representing the ratio between step length and step frequency of an individual, when applied to virtual humans. To this end, we recorded 4 actors (2 males, 2 females) walking at different freely chosen speeds, as well as at different combinations of step frequency and step length. We then performed pilot perceptual studies to identify the ability of viewers to detect the range of walk ratios considered as natural and compared it to the walk ratio freely chosen by the actor when performing walks at the same speeds. Our results will provide new considerations to drive the animation of walking virtual characters using the walk ratio as a parameter, which we believe could enable animators to control the speed of characters through simple parameters while retaining the naturalness of the locomotion.