Fast Bilateral Filtering in Color Images
In this research, we propose a fast approximation to the bilateral filter for color images. The filter is based on two ideas. Firstly, the number of colors which occur in a single natural image is limited. We exploit this color sparseness to rewrite the initial non-linear bilateral filter as a number of linear filter operations. Secondly, we impose a statistical prior to the image values which are locally present within the filter window.
(project page+code)
Bandwidth limited object recognition in high resolution imagery
Current cameras produce high resolution images with detailed information on the objects in the scene, which are only appreciated after zooming into these regions. However this images come at a high computation and bandwith cost when processing and sharing them. In fact it is common to reduce the resolution of these images when processing or sending them, losing potentially useful information. We propose an active information seeking model to extract specific detailed information from images while reducing bandwidth requirements by decreasing the number of pixels to be inspected. (project page + dataset)
Invertible Conditional GANs
In this work, we evaluate encoders to inverse the mapping of a cGAN, i.e., mapping a real image into a latent space and a conditional representation. This allows, for example, to reconstruct and modify real images of faces conditioning on arbitrary attributes. Additionally, we evaluate the design of cGANs. The combination of an encoder with a cGAN, which we call Invertible cGAN (IcGAN), enables to re-generate real images with deterministic complex modifications. (project page + software)
Contextual classification
Contextual information is an important cue for image classification. Our first contribution is the “Multiscale Stacked Sequential learning” algorithm, which is a general method for sequential data. It has been applied to several applications of medical image analysis (e.g. 1, 2, 3). Our research culminated in”Stacked Sequential Scale-Space Taylor Context“, published in TPAMI, where the method surpasses several state-of-the-art methods on 4 datasets: MSRC-21, CAMVID, eTRIMS8 and KAIST2.
Unsupervised learning
This project is led by Ph.D. student Adriana Romero. The main aim of the project is to provide new unsupervised learning algorithms that require a simpler training procedure and more descriptive power than Restricted Boltzmann machines.
Multi-Illuminant estimation
We propose a CRF based method for multi-illuminant estimation. In addition we propose a data set with pixelwise ground-truth of both real scenes and laboratroy setting images.
(project page+data set)
Discriminative Color Descriptors
In this work we propose a number of discriminative color descriptors which are optimizing the discriminative power with respect to a classifciation problem. In addition, we propose a set of universal color descriptors which can be used without any prior training on any data set.
(project page + software)
Synthetic image intrinsic image data set
We propose a synthetic image data set for intrinsic image evaluation.
(project page + data set)
Action Recognition in Still Images
We evaluate the usage of color for action recognition in still images.
(project page)
Color object detection
We extend part-based object detection with color information. Results on VOC PASCAL are provided and code is available.
(project page + software)
Cardiac X-ray analysis
This is a broad project that goes from segmentation to registration with CT data. Up to now the most relevant publications are two segmentation methods (1, 2) and a multi-modal registration algorithm (3). Another research line is devoted to quantification of myocardial staining and detection of diaphragm border (4). This latter is led by Ph.D. student Simeon Petkov.
Compact multi-cue vocabularies
We propose a novel approach for constructing multi-cue Portmanteau vocabularies for image classification.
(project page + software)
Object Recoloring based on Intrinsic Image Estimation
In this research we decompose the image into its intrinsic reflectance components with the aim to recolor scenes.
(project page + software)
Discriminative Pyramids for Object and Scene Recognition
In this research we address the high dimenssionality of spatial pyramids, which is generally considered to be its most serious disadvantage.
(project page + software)
Physics-based color image segmentation
Based on an analysis of the bi-directional reflection model we propose a method which is particularly suited for segmentation in the presence of shadow and highlight edges.
(project page + software)
Color attention for object recognition
We propose a novel image representation where color attention is used to sample the shape description of the image.
(project page + software)
Color Feature Detection for Object Recognition
Luminance edges are still the main source of information in the state-of-the-art methods for feature detection. We propose to exploit the statistical structure of luminance and color in natural images to extract the most discriminative features from the viewpoint of information theory for object recognition.
(project page + software)
Keypoint/Region detection
The project is led by Pedro Martins (University of Coimbra, Portugal), which has been a visiting Ph.D. student at the University of Barcelona. There are two research lines: defining a keypoint extraction procedure based on information theory (1) and introducing a structures highlighting procedure prior to the well known MSER region detector (2, 3) in order to provide stable and complementary regions.