Jordi Gonzàlez

Hunor Laczkó, Meysam Madadi, Sergio Escalera et al.

RGB cloth generation has been deeply studied in the related literature, however, 3D garment generation remains an open problem. In this paper, we build a conditional variational autoencoder for 3D garment generation and draping. We propose a pyramid network to add garment details progressively in a canonical space, i.e. unposing and unshaping the garments w.r.t. the body. We study conditioning the network on surface normal UV maps, as an intermediate representation, which is an easier problem to optimize than 3D coordinates. Our results on two public datasets, CLOTH3D and CAPE, show that our model is robust, controllable in terms of detail generation by the use of multi-resolution pyramids, and achieves state-of-the-art results that can highly generalize to unseen garments, poses, and shapes even when training with small amounts of data.

Bonifaz Stuhr, Jürgen Brauer, Bernhard Schick et al.

A common goal of unpaired image-to-image translation is to preserve content consistency between source images and translated images while mimicking the style of the target domain. Due to biases between the datasets of both domains, many methods suffer from inconsistencies caused by the translation process. Most approaches introduced to mitigate these inconsistencies do not constrain the discriminator, leading to an even more ill-posed training setup. Moreover, none of these approaches is designed for larger crop sizes. In this work, we show that masking the inputs of a global discriminator for both domains with a content-based mask is sufficient to reduce content inconsistencies significantly. However, this strategy leads to artifacts that can be traced back to the masking process. To reduce these artifacts, we introduce a local discriminator that operates on pairs of small crops selected with a similarity sampling strategy. Furthermore, we apply this sampling strategy to sample global input crops from the source and target dataset. In addition, we propose feature-attentive denormalization to selectively incorporate content-based statistics into the generator stream. In our experiments, we show that our method achieves state-of-the-art performance in photorealistic sim-to-real translation and weather translation and also performs well in day-to-night translation. Additionally, we propose the cKVD metric, which builds on the sKVD metric and enables the examination of translation quality at the class or category level.

Diego Velazquez, Pau Rodriguez López, Sergio Alonso et al.

This paper presents EarthView, a comprehensive dataset specifically designed for self-supervision on remote sensing data, intended to enhance deep learning applications on Earth monitoring tasks. The dataset spans 15 tera pixels of global remote-sensing data, combining imagery from a diverse range of sources, including NEON, Sentinel, and a novel release of 1m spatial resolution data from Satellogic. Our dataset provides a wide spectrum of image data with varying resolutions, harnessed from different sensors and organized coherently into an accessible HuggingFace dataset in parquet format. This data spans five years, from 2017 to 2022. Accompanying the dataset, we introduce EarthMAE, a tailored Masked Autoencoder, developed to tackle the distinct challenges of remote sensing data. Trained in a self-supervised fashion, EarthMAE effectively processes different data modalities such as hyperspectral, multispectral, topographical data, segmentation maps, and temporal structure. This model helps us show that pre-training on Satellogic data improves performance on downstream tasks. While there is still a gap to fill in MAE for heterogeneous data, we regard this innovative combination of an expansive, diverse dataset and a versatile model adapted for self-supervised learning as a stride forward in deep learning for Earth monitoring.

Parichehr Behjati, Pau Rodriguez, Armin Mehri et al.

Convolutional neural networks are the most successful models in single image super-resolution. Deeper networks, residual connections, and attention mechanisms have further improved their performance. However, these strategies often improve the reconstruction performance at the expense of considerably increasing the computational cost. This paper introduces a new lightweight super-resolution model based on an efficient method for residual feature and attention aggregation. In order to make an efficient use of the residual features, these are hierarchically aggregated into feature banks for posterior usage at the network output. In parallel, a lightweight hierarchical attention mechanism extracts the most relevant features from the network into attention banks for improving the final output and preventing the information loss through the successive operations inside the network. Therefore, the processing is split into two independent paths of computation that can be simultaneously carried out, resulting in a highly efficient and effective model for reconstructing fine details on high-resolution images from their low-resolution counterparts. Our proposed architecture surpasses state-of-the-art performance in several datasets, while maintaining relatively low computation and memory footprint.

Parichehr Behjati, Pau Rodriguez, Armin Mehri et al.

Super-resolution (SR) has achieved great success due to the development of deep convolutional neural networks (CNNs). However, as the depth and width of the networks increase, CNN-based SR methods have been faced with the challenge of computational complexity in practice. Moreover, most of them train a dedicated model for each target resolution, losing generality and increasing memory requirements. To address these limitations we introduce OverNet, a deep but lightweight convolutional network to solve SISR at arbitrary scale factors with a single model. We make the following contributions: first, we introduce a lightweight recursive feature extractor that enforces efficient reuse of information through a novel recursive structure of skip and dense connections. Second, to maximize the performance of the feature extractor we propose a reconstruction module that generates accurate high-resolution images from overscaled feature maps and can be independently used to improve existing architectures. Third, we introduce a multi-scale loss function to achieve generalization across scales. Through extensive experiments, we demonstrate that our network outperforms previous state-of-the-art results in standard benchmarks while using fewer parameters than previous approaches.

Meysam Madadi, Egils Avots, Sergio Escalera et al.

A new approach for 2D to 3D garment retexturing is proposed based on Gaussian mixture models and thin plate splines (TPS). An automatically segmented garment of an individual is matched to a new source garment and rendered, resulting in augmented images in which the target garment has been retextured by using the texture of the source garment. We divide the problem into garment boundary matching based on Gaussian mixture models and then interpolate inner points using surface topology extracted through geodesic paths, which leads to a more realistic result than standard approaches. We evaluated and compared our system quantitatively by mean square error (MSE) and qualitatively using the mean opinion score (MOS), showing the benefits of the proposed methodology on our gathered dataset.

Pau Rodríguez, Josep M. Gonfaus, Guillem Cucurull et al.

We propose a novel attention mechanism to enhance Convolutional Neural Networks for fine-grained recognition. It learns to attend to lower-level feature activations without requiring part annotations and uses these activations to update and rectify the output likelihood distribution. In contrast to other approaches, the proposed mechanism is modular, architecture-independent and efficient both in terms of parameters and computation required. Experiments show that networks augmented with our approach systematically improve their classification accuracy and become more robust to clutter. As a result, Wide Residual Networks augmented with our proposal surpasses the state of the art classification accuracies in CIFAR-10, the Adience gender recognition task, Stanford dogs, and UEC Food-100.

Pau Rodríguez, Miguel A. Bautista, Jordi Gonzàlez et al.

Target encoding plays a central role when learning Convolutional Neural Networks. In this realm, One-hot encoding is the most prevalent strategy due to its simplicity. However, this so widespread encoding schema assumes a flat label space, thus ignoring rich relationships existing among labels that can be exploited during training. In large-scale datasets, data does not span the full label space, but instead lies in a low-dimensional output manifold. Following this observation, we embed the targets into a low-dimensional space, drastically improving convergence speed while preserving accuracy. Our contribution is two fold: (i) We show that random projections of the label space are a valid tool to find such lower dimensional embeddings, boosting dramatically convergence rates at zero computational cost; and (ii) we propose a normalized eigenrepresentation of the class manifold that encodes the targets with minimal information loss, improving the accuracy of random projections encoding while enjoying the same convergence rates. Experiments on CIFAR-100, CUB200-2011, Imagenet, and MIT Places demonstrate that the proposed approach drastically improves convergence speed while reaching very competitive accuracy rates.

Guillem Cucurull, Pau Rodríguez, V. Oguz Yazici et al.

Social media, as a major platform for communication and information exchange, is a rich repository of the opinions and sentiments of 2.3 billion users about a vast spectrum of topics. To sense the whys of certain social user's demands and cultural-driven interests, however, the knowledge embedded in the 1.8 billion pictures which are uploaded daily in public profiles has just started to be exploited since this process has been typically been text-based. Following this trend on visual-based social analysis, we present a novel methodology based on Deep Learning to build a combined image-and-text based personality trait model, trained with images posted together with words found highly correlated to specific personality traits. So the key contribution here is to explore whether OCEAN personality trait modeling can be addressed based on images, here called \emph{Mind{P}ics}, appearing with certain tags with psychological insights. We found that there is a correlation between those posted images and their accompanying texts, which can be successfully modeled using deep neural networks for personality estimation. The experimental results are consistent with previous cyber-psychology results based on texts or images. In addition, classification results on some traits show that some patterns emerge in the set of images corresponding to a specific text, in essence to those representing an abstract concept. These results open new avenues of research for further refining the proposed personality model under the supervision of psychology experts.

Meysam Madadi, Sergio Escalera, Xavier Baro et al.

Despite recent advances in 3D pose estimation of human hands, especially thanks to the advent of CNNs and depth cameras, this task is still far from being solved. This is mainly due to the highly non-linear dynamics of fingers, which make hand model training a challenging task. In this paper, we exploit a novel hierarchical tree-like structured CNN, in which branches are trained to become specialized in predefined subsets of hand joints, called local poses. We further fuse local pose features, extracted from hierarchical CNN branches, to learn higher order dependencies among joints in the final pose by end-to-end training. Lastly, the loss function used is also defined to incorporate appearance and physical constraints about doable hand motion and deformation. Finally, we introduce a non-rigid data augmentation approach to increase the amount of training depth data. Experimental results suggest that feeding a tree-shaped CNN, specialized in local poses, into a fusion network for modeling joints correlations and dependencies, helps to increase the precision of final estimations, outperforming state-of-the-art results on NYU and SyntheticHand datasets.

Umut Güçlü, Yağmur Güçlütürk, Meysam Madadi et al.

Recent years have seen a sharp increase in the number of related yet distinct advances in semantic segmentation. Here, we tackle this problem by leveraging the respective strengths of these advances. That is, we formulate a conditional random field over a four-connected graph as end-to-end trainable convolutional and recurrent networks, and estimate them via an adversarial process. Importantly, our model learns not only unary potentials but also pairwise potentials, while aggregating multi-scale contexts and controlling higher-order inconsistencies. We evaluate our model on two standard benchmark datasets for semantic face segmentation, achieving state-of-the-art results on both of them.

Pau Rodríguez, Jordi Gonzàlez, Guillem Cucurull et al.

Regularization is key for deep learning since it allows training more complex models while keeping lower levels of overfitting. However, the most prevalent regularizations do not leverage all the capacity of the models since they rely on reducing the effective number of parameters. Feature decorrelation is an alternative for using the full capacity of the models but the overfitting reduction margins are too narrow given the overhead it introduces. In this paper, we show that regularizing negatively correlated features is an obstacle for effective decorrelation and present OrthoReg, a novel regularization technique that locally enforces feature orthogonality. As a result, imposing locality constraints in feature decorrelation removes interferences between negatively correlated feature weights, allowing the regularizer to reach higher decorrelation bounds, and reducing the overfitting more effectively. In particular, we show that the models regularized with OrthoReg have higher accuracy bounds even when batch normalization and dropout are present. Moreover, since our regularization is directly performed on the weights, it is especially suitable for fully convolutional neural networks, where the weight space is constant compared to the feature map space. As a result, we are able to reduce the overfitting of state-of-the-art CNNs on CIFAR-10, CIFAR-100, and SVHN.

Wenjuan Gong, Yongzhen Huang, Jordi Gonzalez et al.

Mixture of parts model has been successfully applied to 2D human pose estimation problem either as explicitly trained body part model or as latent variables for the whole human body model. Mixture of parts model usually utilize tree structure for representing relations between body parts. Tree structures facilitate training and referencing of the model but could not deal with double counting problems, which hinder its applications in 3D pose estimation. While most of work targeted to solve these problems tend to modify the tree models or the optimization target. We incorporate other cues from input features. For example, in surveillance environments, human silhouettes can be extracted relative easily although not flawlessly. In this condition, we can combine extracted human blobs with histogram of gradient feature, which is commonly used in mixture of parts model for training body part templates. The method can be easily extend to other candidate features under our generalized framework. We show 2D body part detection results on a public available dataset: HumanEva dataset. Furthermore, a 2D to 3D pose estimator is trained with Gaussian process regression model and 2D body part detections from the proposed method is fed to the estimator, thus 3D poses are predictable given new 2D body part detections. We also show results of 3D pose estimation on HumanEva dataset.