3D Facial Expressions through Analysis-by-Neural-Synthesis

While existing methods for 3D face reconstruction from in-the-wild images excel at recovering the overall face shape, they commonly miss subtle, extreme, asymmetric, or rarely observed expressions. We improve upon these methods with SMIRK (Spatial Modeling for Image-based Reconstruction of Kinesics), which faithfully reconstructs expressive 3D faces from images. We identify two key limitations in existing methods: shortcomings in their self-supervised training formulation, and a lack of expression diversity in the training images. For training, most methods employ differentiable rendering to compare a predicted face mesh with the input image, along with a plethora of additional loss functions. This differentiable rendering loss not only has to provide supervision to optimize for 3D face geometry, camera, albedo, and lighting, which is an ill-posed optimization problem, but the domain gap between rendering and input image further hinders the learning process. Instead, SMIRK replaces the differentiable rendering with a neural rendering module that, given the rendered predicted mesh geometry, and sparsely sampled pixels of the input image, generates a face image. As the neural rendering gets color information from sampled image pixels, supervising with neural rendering-based reconstruction loss can focus solely on the geometry. Further, it enables us to generate images of the input identity with varying expressions while training. These are then utilized as input to the reconstruction model and used as supervision with ground truth geometry. This effectively augments the training data and enhances the generalization for diverse expressions. Our qualitative, quantitative and particularly our perceptual evaluations demonstrate that SMIRK achieves the new state-of-the art performance on accurate expression reconstruction. Project webpage: https://georgeretsi.github.io/smirk/.

Neural Text to Articulate Talk: Deep Text to Audiovisual Speech Synthesis achieving both Auditory and Photo-realism

Recent advances in deep learning for sequential data have given rise to fast and powerful models that produce realistic videos of talking humans. The state of the art in talking face generation focuses mainly on lip-syncing, being conditioned on audio clips. However, having the ability to synthesize talking humans from text transcriptions rather than audio is particularly beneficial for many applications and is expected to receive more and more attention, following the recent breakthroughs in large language models. For that, most methods implement a cascaded 2-stage architecture of a text-to-speech module followed by an audio-driven talking face generator, but this ignores the highly complex interplay between audio and visual streams that occurs during speaking. In this paper, we propose the first, to the best of our knowledge, text-driven audiovisual speech synthesizer that uses Transformers and does not follow a cascaded approach. Our method, which we call NEUral Text to ARticulate Talk (NEUTART), is a talking face generator that uses a joint audiovisual feature space, as well as speech-informed 3D facial reconstructions and a lip-reading loss for visual supervision. The proposed model produces photorealistic talking face videos with human-like articulation and well-synced audiovisual streams. Our experiments on audiovisual datasets as well as in-the-wild videos reveal state-of-the-art generation quality both in terms of objective metrics and human evaluation.

Pre-training Music Classification Models via Music Source Separation

In this paper, we study whether music source separation can be used as a pre-training strategy for music representation learning, targeted at music classification tasks. To this end, we first pre-train U-Net networks under various music source separation objectives, such as the isolation of vocal or instrumental sources from a musical piece; afterwards, we attach a convolutional tail network to the pre-trained U-Net and jointly finetune the whole network. The features learned by the separation network are also propagated to the tail network through skip connections. Experimental results in two widely used and publicly available datasets indicate that pre-training the U-Nets with a music source separation objective can improve performance compared to both training the whole network from scratch and using the tail network as a standalone in two music classification tasks: music auto-tagging, when vocal separation is used, and music genre classification for the case of multi-source separation.

Feather: An Elegant Solution to Effective DNN Sparsification

Neural Network pruning is an increasingly popular way for producing compact and efficient models, suitable for resource-limited environments, while preserving high performance. While the pruning can be performed using a multi-cycle training and fine-tuning process, the recent trend is to encompass the sparsification process during the standard course of training. To this end, we introduce Feather, an efficient sparse training module utilizing the powerful Straight-Through Estimator as its core, coupled with a new thresholding operator and a gradient scaling technique, enabling robust, out-of-the-box sparsification performance. Feather's effectiveness and adaptability is demonstrated using various architectures on the CIFAR dataset, while on ImageNet it achieves state-of-the-art Top-1 validation accuracy using the ResNet-50 architecture, surpassing existing methods, including more complex and computationally heavy ones, by a considerable margin. Code is publicly available at https://github.com/athglentis/feather .

Matrix Factorization in Tropical and Mixed Tropical-Linear Algebras

Matrix Factorization (MF) has found numerous applications in Machine Learning and Data Mining, including collaborative filtering recommendation systems, dimensionality reduction, data visualization, and community detection. Motivated by the recent successes of tropical algebra and geometry in machine learning, we investigate two problems involving matrix factorization over the tropical algebra. For the first problem, Tropical Matrix Factorization (TMF), which has been studied already in the literature, we propose an improved algorithm that avoids many of the local optima. The second formulation considers the approximate decomposition of a given matrix into the product of three matrices where a usual matrix product is followed by a tropical product. This formulation has a very interesting interpretation in terms of the learning of the utility functions of multiple users. We also present numerical results illustrating the effectiveness of the proposed algorithms, as well as an application to recommendation systems with promising results.

Photorealistic and Identity-Preserving Image-Based Emotion Manipulation with Latent Diffusion Models

In this paper, we investigate the emotion manipulation capabilities of diffusion models with "in-the-wild" images, a rather unexplored application area relative to the vast and rapidly growing literature for image-to-image translation tasks. Our proposed method encapsulates several pieces of prior work, with the most important being Latent Diffusion models and text-driven manipulation with CLIP latents. We conduct extensive qualitative and quantitative evaluations on AffectNet, demonstrating the superiority of our approach in terms of image quality and realism, while achieving competitive results relative to emotion translation compared to a variety of GAN-based counterparts. Code is released as a publicly available repo.

Revisiting Tropical Polynomial Division: Theory, Algorithms and Application to Neural Networks

Tropical geometry has recently found several applications in the analysis of neural networks with piecewise linear activation functions. This paper presents a new look at the problem of tropical polynomial division and its application to the simplification of neural networks. We analyze tropical polynomials with real coefficients, extending earlier ideas and methods developed for polynomials with integer coefficients. We first prove the existence of a unique quotient-remainder pair and characterize the quotient in terms of the convex bi-conjugate of a related function. Interestingly, the quotient of tropical polynomials with integer coefficients does not necessarily have integer coefficients. Furthermore, we develop a relationship of tropical polynomial division with the computation of the convex hull of unions of convex polyhedra and use it to derive an exact algorithm for tropical polynomial division. An approximate algorithm is also presented, based on an alternation between data partition and linear programming. We also develop special techniques to divide composite polynomials, described as sums or maxima of simpler ones. Finally, we present some numerical results to illustrate the efficiency of the algorithms proposed, using the MNIST handwritten digit and CIFAR-10 datasets.

ViDaS Video Depth-aware Saliency Network

We introduce ViDaS, a two-stream, fully convolutional Video, Depth-Aware Saliency network to address the problem of attention modeling ``in-the-wild", via saliency prediction in videos. Contrary to existing visual saliency approaches using only RGB frames as input, our network employs also depth as an additional modality. The network consists of two visual streams, one for the RGB frames, and one for the depth frames. Both streams follow an encoder-decoder approach and are fused to obtain a final saliency map. The network is trained end-to-end and is evaluated in a variety of different databases with eye-tracking data, containing a wide range of video content. Although the publicly available datasets do not contain depth, we estimate it using three different state-of-the-art methods, to enable comparisons and a deeper insight. Our method outperforms in most cases state-of-the-art models and our RGB-only variant, which indicates that depth can be beneficial to accurately estimating saliency in videos displayed on a 2D screen. Depth has been widely used to assist salient object detection problems, where it has been proven to be very beneficial. Our problem though differs significantly from salient object detection, since it is not restricted to specific salient objects, but predicts human attention in a more general aspect. These two problems not only have different objectives, but also different ground truth data and evaluation metrics. To our best knowledge, this is the first competitive deep learning video saliency estimation approach that combines both RGB and Depth features to address the general problem of saliency estimation ``in-the-wild". The code will be publicly released.

OVeNet: Offset Vector Network for Semantic Segmentation

Semantic segmentation is a fundamental task in visual scene understanding. We focus on the supervised setting, where ground-truth semantic annotations are available. Based on knowledge about the high regularity of real-world scenes, we propose a method for improving class predictions by learning to selectively exploit information from neighboring pixels. In particular, our method is based on the prior that for each pixel, there is a seed pixel in its close neighborhood sharing the same prediction with the former. Motivated by this prior, we design a novel two-head network, named Offset Vector Network (OVeNet), which generates both standard semantic predictions and a dense 2D offset vector field indicating the offset from each pixel to the respective seed pixel, which is used to compute an alternative, seed-based semantic prediction. The two predictions are adaptively fused at each pixel using a learnt dense confidence map for the predicted offset vector field. We supervise offset vectors indirectly via optimizing the seed-based prediction and via a novel loss on the confidence map. Compared to the baseline state-of-the-art architectures HRNet and HRNet+OCR on which OVeNet is built, the latter achieves significant performance gains on two prominent benchmarks for semantic segmentation of driving scenes, namely Cityscapes and ACDC. Code is available at https://github.com/stamatisalex/OVeNet

Medical Face Masks and Emotion Recognition from the Body: Insights from a Deep Learning Perspective

The COVID-19 pandemic has undoubtedly changed the standards and affected all aspects of our lives, especially social life. It has forced people to extensively wear medical face masks, in order to prevent transmission. This face occlusion can strongly irritate emotional reading from the face and urges us to incorporate the whole body for emotion recognition, as it needs to play a more major role, despite its complementary nature. In this paper, we want to conduct insightful studies about the effect of face occlusion on emotion recognition performance, and showcase the superiority of full body input over plain masked face. We utilize a deep learning model based on the Temporal Segment Network framework and aspire to fully overcome the consequences of the face mask. Although single RGB stream models can adapt and learn both facial and bodily features, this may lead to irrelevant information confusion. By processing those features separately and fusing their preliminary prediction scores with a late fusion scheme, we are more effectively taking advantage of both modalities. This architecture can also naturally support temporal modeling, by mingling information among neighboring segment frames. Experimental results suggest that spatial structure plays a more important role for an emotional expression, while temporal structure is complementary.