Attila Szabó

Attila Szabo, Hadi Jamali-Rad, Siva-Datta Mannava

Traditional empirical risk minimization (ERM) for semantic segmentation can disproportionately advantage or disadvantage certain target classes in favor of an (unfair but) improved overall performance. Inspired by the recently introduced tilted ERM (TERM), we propose tilted cross-entropy (TCE) loss and adapt it to the semantic segmentation setting to minimize performance disparity among target classes and promote fairness. Through quantitative and qualitative performance analyses, we demonstrate that the proposed Stochastic TCE for semantic segmentation can efficiently improve the low-performing classes of Cityscapes and ADE20k datasets trained with multi-class cross-entropy (MCCE), and also results in improved overall fairness.

Hadi Jamali-Rad, Attila Szabo

Semantic segmentation is one of the most fundamental problems in computer vision with significant impact on a wide variety of applications. Adversarial learning is shown to be an effective approach for improving semantic segmentation quality by enforcing higher-level pixel correlations and structural information. However, state-of-the-art semantic segmentation models cannot be easily plugged into an adversarial setting because they are not designed to accommodate convergence and stability issues in adversarial networks. We bridge this gap by building a conditional adversarial network with a state-of-the-art segmentation model (DeepLabv3+) at its core. To battle the stability issues, we introduce a novel lookahead adversarial learning (LoAd) approach with an embedded label map aggregation module. We focus on semantic segmentation models that run fast at inference for near real-time field applications. Through extensive experimentation, we demonstrate that the proposed solution can alleviate divergence issues in an adversarial semantic segmentation setting and results in considerable performance improvements (+5% in some classes) on the baseline for three standard datasets.

Attila Szabó, Givi Meishvili, Paolo Favaro

In this paper we present, to the best of our knowledge, the first method to learn a generative model of 3D shapes from natural images in a fully unsupervised way. For example, we do not use any ground truth 3D or 2D annotations, stereo video, and ego-motion during the training. Our approach follows the general strategy of Generative Adversarial Networks, where an image generator network learns to create image samples that are realistic enough to fool a discriminator network into believing that they are natural images. In contrast, in our approach the image generation is split into 2 stages. In the first stage a generator network outputs 3D objects. In the second, a differentiable renderer produces an image of the 3D objects from random viewpoints. The key observation is that a realistic 3D object should yield a realistic rendering from any plausible viewpoint. Thus, by randomizing the choice of the viewpoint our proposed training forces the generator network to learn an interpretable 3D representation disentangled from the viewpoint. In this work, a 3D representation consists of a triangle mesh and a texture map that is used to color the triangle surface by using the UV-mapping technique. We provide analysis of our learning approach, expose its ambiguities and show how to overcome them. Experimentally, we demonstrate that our method can learn realistic 3D shapes of faces by using only the natural images of the FFHQ dataset.

Attila Szabó, Paolo Favaro

We present a method to learn the 3D surface of objects directly from a collection of images. Previous work achieved this capability by exploiting additional manual annotation, such as object pose, 3D surface templates, temporal continuity of videos, manually selected landmarks, and foreground/background masks. In contrast, our method does not make use of any such annotation. Rather, it builds a generative model, a convolutional neural network, which, given a noise vector sample, outputs the 3D surface and texture of an object and a background image. These 3 components combined with an additional random viewpoint vector are then fed to a differential renderer to produce a view of the sampled object and background. Our general principle is that if the output of the renderer, the generated image, is realistic, then its input, the generated 3D and texture, should also be realistic. To achieve realism, the generative model is trained adversarially against a discriminator that tries to distinguish between the output of the renderer and real images from the given data set. Moreover, our generative model can be paired with an encoder and trained as an autoencoder, to automatically extract the 3D shape, texture and pose of the object in an image. Our trained generative model and encoder show promising results both on real and synthetic data, which demonstrate for the first time that fully unsupervised 3D learning from image collections is possible.

Tiziano Portenier, Qiyang Hu, Attila Szabó et al.

We present a novel system for sketch-based face image editing, enabling users to edit images intuitively by sketching a few strokes on a region of interest. Our interface features tools to express a desired image manipulation by providing both geometry and color constraints as user-drawn strokes. As an alternative to the direct user input, our proposed system naturally supports a copy-paste mode, which allows users to edit a given image region by using parts of another exemplar image without the need of hand-drawn sketching at all. The proposed interface runs in real-time and facilitates an interactive and iterative workflow to quickly express the intended edits. Our system is based on a novel sketch domain and a convolutional neural network trained end-to-end to automatically learn to render image regions corresponding to the input strokes. To achieve high quality and semantically consistent results we train our neural network on two simultaneous tasks, namely image completion and image translation. To the best of our knowledge, we are the first to combine these two tasks in a unified framework for interactive image editing. Our results show that the proposed sketch domain, network architecture, and training procedure generalize well to real user input and enable high quality synthesis results without additional post-processing.

Qiyang Hu, Attila Szabó, Tiziano Portenier et al.

We propose an approach to learn image representations that consist of disentangled factors of variation without exploiting any manual labeling or data domain knowledge. A factor of variation corresponds to an image attribute that can be discerned consistently across a set of images, such as the pose or color of objects. Our disentangled representation consists of a concatenation of feature chunks, each chunk representing a factor of variation. It supports applications such as transferring attributes from one image to another, by simply mixing and unmixing feature chunks, and classification or retrieval based on one or several attributes, by considering a user-specified subset of feature chunks. We learn our representation without any labeling or knowledge of the data domain, using an autoencoder architecture with two novel training objectives: first, we propose an invariance objective to encourage that encoding of each attribute, and decoding of each chunk, are invariant to changes in other attributes and chunks, respectively; second, we include a classification objective, which ensures that each chunk corresponds to a consistently discernible attribute in the represented image, hence avoiding degenerate feature mappings where some chunks are completely ignored. We demonstrate the effectiveness of our approach on the MNIST, Sprites, and CelebA datasets.

Attila Szabó, Qiyang Hu, Tiziano Portenier et al.

We study the problem of building models that disentangle independent factors of variation. Such models could be used to encode features that can efficiently be used for classification and to transfer attributes between different images in image synthesis. As data we use a weakly labeled training set. Our weak labels indicate what single factor has changed between two data samples, although the relative value of the change is unknown. This labeling is of particular interest as it may be readily available without annotation costs. To make use of weak labels we introduce an autoencoder model and train it through constraints on image pairs and triplets. We formally prove that without additional knowledge there is no guarantee that two images with the same factor of variation will be mapped to the same feature. We call this issue the reference ambiguity. Moreover, we show the role of the feature dimensionality and adversarial training. We demonstrate experimentally that the proposed model can successfully transfer attributes on several datasets, but show also cases when the reference ambiguity occurs.