Abstract: This paper investigates the role of saliency to improve the classification accuracy of a Convolutional Neural Network (CNN) for the case when scarce training data is available. Our approach consists in adding a saliency branch to an existing CNN architecture which is used to modulate the standard bottom-up visual features from the original image input, acting as an attentional mechanism that guides the feature extraction process. The main aim of the proposed approach is to enable the effective training of a fine-grained recognition model with limited training samples and to improve the performance on the task, thereby alleviating the need to annotate a large dataset. The vast majority of saliency methods are evaluated on their ability to generate saliency maps, and not on their functionality in a complete vision pipeline. Our proposed pipeline allows to evaluate saliency methods for the high-level task of object recognition. We perform extensive experiments on various fine-grained datasets (Flowers, Birds, Cars, and Dogs) under different conditions and show that saliency can considerably improve the network’s performance, especially for the case of scarce training data. Furthermore, our experiments show that saliency methods that obtain improved saliency maps (as measured by traditional saliency benchmarks) also translate to saliency methods that yield improved performance gains when applied in an object recognition pipeline.

Abstract: In this study we provide the analysis of eye movement behavior elicited by low-level feature distinctiveness with a dataset of synthetically-generated image patterns. Design of visual stimuli was inspired by the ones used in previous psychophysical experiments, namely in free-viewing and visual searching tasks, to provide a total of 15 types of stimuli, divided according to the task and feature to be analyzed. Our interest is to analyze the influences of low-level feature contrast between a salient region and the rest of distractors, providing fixation localization characteristics and reaction time of landing inside the salient region. Eye-tracking data was collected from 34 participants during the viewing of a 230 images dataset. Results show that saliency is predominantly and distinctively influenced by: 1. feature type, 2. feature contrast, 3. temporality of fixations, 4. task difficulty and 5. center bias. This experimentation proposes a new psychophysical basis for saliency model evaluation using synthetic images.

Abstract: Deep video action recognition models have been highly successful in recent years but require large quantities of manually-annotated data, which are expensive and laborious to obtain. In this work, we investigate the generation of synthetic training data for video action recognition, as synthetic data have been successfully used to supervise models for a variety of other computer vision tasks. We propose an interpretable parametric generative model of human action videos that relies on procedural generation, physics models and other components of modern game engines. With this model we generate a diverse, realistic, and physically plausible dataset of human action videos, called PHAV for “Procedural Human Action Videos”. PHAV contains a total of 39,982 videos, with more than 1000 examples for each of 35 action categories. Our video generation approach is not limited to existing motion capture sequences: 14 of these 35 categories are procedurally-defined synthetic actions. In addition, each video is represented with 6 different data modalities, including RGB, optical flow and pixel-level semantic labels. These modalities are generated almost simultaneously using the Multiple Render Targets feature of modern GPUs. In order to leverage PHAV, we introduce a deep multi-task (i.e. that considers action classes from multiple datasets) representation learning architecture that is able to simultaneously learn from synthetic and real video datasets, even when their action categories differ. Our experiments on the UCF-101 and HMDB-51 benchmarks suggest that combining our large set of synthetic videos with small real-world datasets can boost recognition performance. Our approach also significantly outperforms video representations produced by fine-tuning state-of-the-art unsupervised generative models of videos.

Abstract: The impressive performance of Convolutional Neural Networks (CNNs) when solving different vision problems is shadowed by their black-box nature and our consequent lack of understanding of the representations they build and how these representations are organized. To help understanding these issues, we propose to describe the activity of individual neurons by their Neuron Feature visualization and quantify their inherent selectivity with two specific properties. We explore selectivity indexes for: an image feature (color); and an image label (class membership). Our contribution is a framework to seek or classify neurons by indexing on these selectivity properties. It helps to find color selective neurons, such as a red-mushroom neuron in layer Conv4 or class selective neurons such as dog-face neurons in layer Conv5 in VGG-M, and establishes a methodology to derive other selectivity properties. Indexing on neuron selectivity can statistically draw how features and classes are represented through layers in a moment when the size of trained nets is growing and automatic tools to index neurons can be helpful.

Abstract: Approaches that use more than two consecutive video frames in the optical flow estimation have a long research history. However, almost all such methods utilize extra information for a pre-processing flow prediction or for a post-processing flow correction and filtering. In contrast, this paper differs from previously developed techniques. We propose a new algorithm for the likelihood function calculation (alternatively the matching cost volume) that is used in the maximum a posteriori estimation. We exploit the fact that in general, optical flow is locally constant in the sense of time and the likelihood function depends on both the previous and the future frame. Implementation of our idea increases the robustness of optical flow estimation. As a result, our method outperforms 9% over the DCFlow technique, which we use as prototype for our CNN based computation architecture, on the most challenging MPI-Sintel dataset for the non-occluded mask metric. Furthermore, our approach considerably increases the accuracy of the flow estimation for the matching cost processing, consequently outperforming the original DCFlow algorithm results up to 50% in occluded regions and up to 9% in non-occluded regions on the MPI-Sintel dataset. The experimental section shows that the proposed method achieves state-of-the-arts results especially on the MPI-Sintel dataset.

Abstract: In this work, we propose a convolutional neural network based approach to estimate the spectral reflectance of a surface and spectral power distribution of light from a single RGB image of a V-shaped surface. Interreflections happening in a concave surface lead to gradients of RGB values over its area. These gradients carry a lot of information concerning the physical properties of the surface and the illuminant. Our network is trained with only simulated data constructed using a physics-based interreflection model. Coupling interreflection effects with deep learning helps to retrieve the spectral reflectance under an unknown light and to estimate spectral power distribution of this light as well. In addition, it is more robust to the presence of image noise than classical approaches. Our results show that the proposed approach outperforms state-of-the-art learning-based approaches on simulated data. In addition, it gives better results on real data compared to other interreflection-based approaches.

Abstract: This paper addresses the problem of inferring unseen cross-modal image-to-image translations between multiple modalities. We assume that only some of the pairwise translations have been seen (i.e. trained) and infer the remaining unseen translations (where training pairs are not available). We propose mix and match networks, an approach where multiple encoders and decoders are aligned in such a way that the desired translation can be obtained by simply cascading the source encoder and the target decoder, even when they have not interacted during the training stage (i.e. unseen). The main challenge lies in the alignment of the latent representations at the bottlenecks of encoder-decoder pairs. We propose an architecture with several tools to encourage alignment, including autoencoders and robust side information and latent consistency losses. We show the benefits of our approach in terms of effectiveness and scalability compared with other pairwise image-to-image translation approaches. We also propose zero-pair cross-modal image translation, a challenging setting where the objective is inferring semantic segmentation from depth (and vice-versa) without explicit segmentation-depth pairs, and only from two (disjoint) segmentation-RGB and depth-RGB training sets. We observe that a certain part of the shared information between unseen modalities might not be reachable, so we further propose a variant that leverages pseudo-pairs which allows us to exploit this shared information between the unseen modalities

Abstract: JCR 2019 Q2, IF=3.121
The task of unpaired image-to-image translation is highly challenging due to the lack of explicit cross-domain pairs of instances. We consider here diverse image translation (DIT), an even more challenging setting in which an image can have multiple plausible translations. This is normally achieved by explicitly disentangling content and style in the latent representation and sampling different styles codes while maintaining the image content. Despite the success of current DIT models, they are prone to suffer from bias. In this paper, we study the problem of bias in image-to-image translation. Biased datasets may add undesired changes (e.g. change gender or race in face images) to the output translations as a consequence of the particular underlying visual distribution in the target domain. In order to alleviate the effects of this problem we propose the use of semantic constraints that enforce the preservation of desired image properties. Our proposed model is a step towards unbiased diverse image-to-image translation (UDIT), and results in less unwanted changes in the translated images while still performing the wanted transformation. Experiments on several heavily biased datasets show the effectiveness of the proposed techniques in different domains such as faces, objects, and scenes.

Abstract: Salient object segmentation in videos is generally broken up in a video segmentation part and a saliency assignment part. Recently, object proposals, which are used to segment the image, have had significant impact on many computer vision applications, including image segmentation, object detection, and recently saliency detection in still images. However, their usage has not yet been evaluated for salient object segmentation in videos. Therefore, in this paper, we investigate the application of object proposals to salient object segmentation in videos. In addition, we propose a new motion feature derived from the optical flow structure tensor for video saliency detection. Experiments on two standard benchmark datasets for video saliency show that the proposed motion feature improves saliency estimation results, and that object proposals are an efficient method for salient object segmentation. Results on the challenging SegTrack v2 and Fukuchi benchmark data sets show that we significantly outperform the state-of-the-art.