Abstract: Deep Neural Networks trained on large datasets can be easily transferred to new domains with far fewer labeled examples by a process called fine-tuning. This has the advantage that representations learned in the large source domain can be exploited on smaller target domains. However, networks designed to be optimal for the source task are often prohibitively large for the target task. In this work we address the compression of networks after domain transfer.
We focus on compression algorithms based on low-rank matrix decomposition. Existing methods base compression solely on learned network weights and ignore the statistics of network activations. We show that domain transfer leads to large shifts in network activations and that it is desirable to take this into account when compressing.
We demonstrate that considering activation statistics when compressing weights leads to a rank-constrained regression problem with a closed-form solution. Because our method takes into account the target domain, it can more optimally
remove the redundancy in the weights. Experiments show that our Domain Adaptive Low Rank (DALR) method significantly outperforms existing low-rank compression techniques. With our approach, the fc6 layer of VGG19 can be compressed more than 4x more than using truncated SVD alone – with only a minor or no loss in accuracy. When applied to domain-transferred networks it allows for compression down to only 5-20% of the original number of parameters with only a minor drop in performance.

Abstract: Deep convolutional architectures have shown to be successful frameworks to solve generic computer vision problems. The estimation of intrinsic reflectance from single image is not a solved problem yet. Encoder-Decoder architectures are a perfect approach for pixel-wise reflectance estimation, although it usually suffers from the lack of large datasets. Lack of data can be partially solved with data augmentation, however usual techniques focus on geometric changes which does not help for reflectance estimation. In this paper we propose a color-based data augmentation technique that extends the training data by increasing the variability of chromaticity. Rotation on the red-green blue-yellow plane of an opponent space enable to increase the training set in a coherent and sound way that improves network generalization capability for reflectance estimation. We perform some experiments on the Sintel dataset showing that our color-based augmentation increase performance and overcomes one of the state-of-the-art methods.

Abstract: Convolutional Neural Networks have been proposed as suitable frameworks to model biological vision. Some of these artificial networks showed representational properties that rival primate performances in object recognition. In this paper we explore how color is encoded in a trained artificial network. It is performed by estimating a color selectivity index for each neuron, which allows us to describe the neuron activity to a color input stimuli. The index allows us to classify whether they are color selective or not and if they are of a single or double color. We have determined that all five convolutional layers of the network have a large number of color selective neurons. Color opponency clearly emerges in the first layer, presenting 4 main axes (Black-White, Red-Cyan, Blue-Yellow and Magenta-Green), but this is reduced and rotated as we go deeper into the network. In layer 2 we find a denser hue sampling of color neurons and opponency is reduced almost to one new main axis, the Bluish-Orangish coinciding with the dataset bias. In layers 3, 4 and 5 color neurons are similar amongst themselves, presenting different type of neurons that detect specific colored objects (e.g., orangish faces), specific surrounds (e.g., blue sky) or specific colored or contrasted object-surround configurations (e.g. blue blob in a green surround). Overall, our work concludes that color and shape representation are successively entangled through all the layers of the studied network, revealing certain parallelisms with the reported evidences in primate brains that can provide useful insight into intermediate hierarchical spatio-chromatic representations.

Abstract: Conditional image synthesis is an integral part of many X2I translation systems, including image-to-image, text-to-image and audio-to-image translation systems. Training these large systems generally requires huge amounts of training data.
Therefore, we investigate knowledge distillation to transfer knowledge from a high-quality unconditioned generative model (e.g., StyleGAN) to a conditioned synthetic image generation modules in a variety of systems. To initialize the conditional and reference branch (from a unconditional GAN) we exploit the style mixing characteristics of high-quality GANs to generate an infinite supply of style-mixed triplets to perform the knowledge distillation. Extensive experimental results in a number of image generation tasks (i.e., image-to-image, semantic segmentation-to-image, text-to-image and audio-to-image) demonstrate qualitatively and quantitatively that our method successfully transfers knowledge to the synthetic image generation modules, resulting in more realistic images than previous methods as confirmed by a significant drop in the FID.

Abstract: Computer vision has established a foothold in the online fashion retail industry. Main product detection is a crucial step of vision-based fashion product feed parsing pipelines, focused on identifying the bounding boxes that contain the product being sold in the gallery of images of the product page. The current state-of-the-art approach does not leverage the relations between regions in the image, and treats images of the same product independently, therefore not fully exploiting visual and product contextual information. In this paper, we propose a model that incorporates Graph Convolutional Networks (GCN) that jointly represent all detected bounding boxes in the gallery as nodes. We show that the proposed method is better than the state-of-the-art, especially, when we consider the scenario where title-input is missing at inference time and for cross-dataset evaluation, our method outperforms previous approaches by a large margin.

Abstract: Computational color constancy is a fundamental prerequisite for many computer vision applications. This paper presents a survey of many recent developments and state-of-the- art methods. Several criteria are proposed that are used to assess the approaches. A taxonomy of existing algorithms is proposed and methods are separated in three groups: static methods, gamut-based methods and learning-based methods. Further, the experimental setup is discussed including an overview of publicly available data sets. Finally, various freely available methods, of which some are considered to be state-of-the-art, are evaluated on two data sets.

Abstract: Comic books constitute an important cultural heritage asset in many countries. Digitization combined with subsequent comic book understanding would enable a variety of new applications, including content-based retrieval and content retargeting. Document understanding in this domain is challenging as comics are semi-structured documents, combining semantically important graphical and textual parts. Few studies have been done in this direction. In this work we detail a novel approach for closed and non-closed speech balloon localization in scanned comic book pages, an essential step towards a fully automatic comic book understanding. The approach is compared with existing methods for closed balloon localization found in the literature and results are presented.

Abstract: Colour is derived from three physical properties: incident light, object reflectance and sensor sensitivities. Incident light varies under natural conditions; hence, recovering scene illuminant is an important issue in computational colour. One way to deal with this problem under calibrated conditions is by following three steps, 1) building a narrow-band sensor basis to accomplish the diagonal model, 2) building a feasible set of illuminants, and 3) defining criteria to select the best illuminant. In this work we focus on colour constancy for natural images by introducing perceptual criteria in the first and third stages.
To deal with the illuminant selection step, we hypothesise that basic colour categories can be used as anchor categories to recover the best illuminant. These colour names are related to the way that the human visual system has evolved to encode relevant natural colour statistics. Therefore the recovered image provides the best representation of the scene labelled with the basic colour terms. We demonstrate with several experiments how this selection criterion achieves current state-of-art results in computational colour constancy. In addition to this result, we psychophysically prove that usual angular error used in colour constancy does not correlate with human preferences, and we propose a new perceptual colour constancy evaluation.
The implementation of this selection criterion strongly relies on the use of a diagonal
model for illuminant change. Consequently, the second contribution focuses on building an appropriate narrow-band sensor basis to represent natural images. We propose to use the spectral sharpening technique to compute a unique narrow-band basis optimised to represent a large set of natural reflectances under natural illuminants and given in the basis of human cones. The proposed sensors allow predicting unique hues and the World colour Survey data independently of the illuminant by using a compact singularity function. Additionally, we studied different families of sharp sensors to minimise different perceptual measures. This study brought us to extend the spherical sampling procedure from 3D to 6D.
Several research lines still remain open. One natural extension would be to measure the
effects of using the computed sharp sensors on the category hypothesis, while another might be to insert spatial contextual information to improve category hypothesis. Finally, much work still needs to be done to explore how individual sensors can be adjusted to the colours in a scene.

Abstract: Color names are required in real-world applications such as image retrieval and image annotation. Traditionally, they are learned from a collection of labelled color chips. These color chips are labelled with color names within a well-defined experimental setup by human test subjects. However naming colors in real-world images differs significantly from this experimental setting. In this paper, we investigate how color names learned from color chips compare to color names learned from real-world images. To avoid hand labelling real-world images with color names we use Google Image to collect a data set. Due to limitations of Google Image this data set contains a substantial quantity of wrongly labelled data. We propose several variants of the PLSA model to learn color names from this noisy data. Experimental results show that color names learned from real-world images significantly outperform color names learned from labelled color chips for both image retrieval and image annotation.