How to estimate carbon footprint when training deep learning models? A guide and review.

Machine learning and deep learning models have become essential in the recent fast development of artificial intelligence in many sectors of the society. It is now widely acknowledge that the development of these models has an environmental cost that has been analyzed in many studies. Several online and software tools have been developed to track energy consumption while training machine learning models. In this paper, we propose a comprehensive introduction and comparison of these tools for AI practitioners wishing to start estimating the environmental impact of their work. We review the specific vocabulary, the technical requirements for each tool. We compare the energy consumption estimated by each tool on two deep neural networks for image processing and on different types of servers. From these experiments, we provide some advice for better choosing the right tool and infrastructure.

Multi-Task Deep Learning for Image Segmentation Using Recursive Approximation Tasks.

Fully supervised deep neural networks for segmentation usually require a massive amount of pixel-level labels which are manually expensive to create. In this work, we develop a multi-task learning method to relax this constraint. We regard the segmentation problem as a sequence of approximation subproblems that are recursively defined and in increasing levels of approximation accuracy. The subproblems are handled by a framework that consists of 1) a segmentation task that learns from pixel-level ground truth segmentation masks of a small fraction of the images, 2) a recursive approximation task that conducts partial object regions learning and data-driven mask evolution starting from partial masks of each object instance, and 3) other problem oriented auxiliary tasks that are trained with sparse annotations and promote the learning of dedicated features. Most training images are only labeled by (rough) partial masks, which do not contain exact object boundaries, rather than by their full segmentation masks. During the training phase, the approximation task learns the statistics of these partial masks, and the partial regions are recursively increased towards object boundaries aided by the learned information from the segmentation task in a fully data-driven fashion. The network is trained on an extremely small amount of precisely segmented images and a large set of coarse labels. Annotations can thus be obtained in a cheap way. We demonstrate the efficiency of our approach in three applications with microscopy images and ultrasound images.

Variational Osmosis for Non-linear Image Fusion.

We propose a new variational model for non-linear image fusion. Our approach is based on the use of an osmosis energy term related to the one studied in Vogel et al. [44] and Weickert et al. [45]. The minimization of the proposed non-convex energy realizes visually plausible image data fusion, invariant to multiplicative brightness changes. On the practical side, it requires minimal supervision and parameter tuning and can encode prior information on the structure of the images to be fused. For the numerical solution of the proposed model, we develop a primal-dual algorithm and we apply the resulting minimization scheme to solve multi-modal face fusion, color transfer and cultural heritage conservation problems. Visual and quantitative comparisons to state-of-the-art approaches prove the out-performance and the flexibility of our method.

SuperPatchMatch: An Algorithm for Robust Correspondences Using Superpixel Patches.

Superpixels have become very popular in many computer vision applications. Nevertheless, they remain under-exploited, since the superpixel decomposition may produce irregular and nonstable segmentation results due to the dependency to the image content. In this paper, we first introduce a novel structure, a superpixel-based patch, called SuperPatch. The proposed structure, based on superpixel neighborhood, leads to a robust descriptor, since spatial information is naturally included. The generalization of the PatchMatch method to SuperPatches, named SuperPatchMatch, is introduced. Finally, we propose a framework to perform fast segmentation and labeling from an image database, and demonstrate the potential of our approach, since we outperform, in terms of computational cost and accuracy, the results of state-of-the-art methods on both face labeling and medical image segmentation.

Variational exemplar-based image colorization.

In this paper, we address the problem of recovering a color image from a grayscale one. The input color data comes from a source image considered as a reference image. Reconstructing the missing color of a grayscale pixel is here viewed as the problem of automatically selecting the best color among a set of color candidates while simultaneously ensuring the local spatial coherency of the reconstructed color information. To solve this problem, we propose a variational approach where a specific energy is designed to model the color selection and the spatial constraint problems simultaneously. The contributions of this paper are twofold. First, we introduce a variational formulation modeling the color selection problem under spatial constraints and propose a minimization scheme, which computes a local minima of the defined nonconvex energy. Second, we combine different patch-based features and distances in order to construct a consistent set of possible color candidates. This set is used as input data and our energy minimization automatically selectsthe best color to transfer for each pixel of the grayscale image. Finally, the experiments illustrate the potentiality of our simple methodology and show that our results are very competitive with respect to the state-of-the-art methods.

Image editing with spatiograms transfer.

Histogram equalization is a well-known method for image contrast enhancement. Nevertheless, as histograms do not include any information on the spatial repartition of colors, their application to local image editing problems remains limited. To cope with this lack of spatial information, spatiograms have been recently proposed for tracking purposes. A spatiogram is an image descriptor that combines a histogram with the mean and the variance of the position of each color. In this paper, we address the problem of local retouching of images by proposing a variational method for spatiogram transfer. More precisely, a reference spatiogram is used to modify the color value of a given region of interest of the processed image. Experiments on shadow removal and inpainting demonstrate the strength of the proposed approach.

Tracking with occlusions via graph cuts.

This work presents a new method for tracking and segmenting along time-interacting objects within an image sequence. One major contribution of the paper is the formalization of the notion of visible and occluded parts. For each object, we aim at tracking these two parts. Assuming that the velocity of each object is driven by a dynamical law, predictions can be used to guide the successive estimations. Separating these predicted areas into good and bad parts with respect to the final segmentation and representing the objects with their visible and occluded parts permit handling partial and complete occlusions. To achieve this tracking, a label is assigned to each object and an energy function representing the multilabel problem is minimized via a graph cuts optimization. This energy contains terms based on image intensities which enable segmenting and regularizing the visible parts of the objects. It also includes terms dedicated to the management of the occluded and disappearing areas, which are defined on the areas of prediction of the objects. The results on several challenging sequences prove the strength of the proposed approach.

A comprehensive framework for image inpainting.

Inpainting is the art of modifying an image in a form that is not detectable by an ordinary observer. There are numerous and very different approaches to tackle the inpainting problem, though as explained in this paper, the most successful algorithms are based upon one or two of the following three basic techniques: copy-and-paste texture synthesis, geometric partial differential equations (PDEs), and coherence among neighboring pixels. We combine these three building blocks in a variational model, and provide a working algorithm for image inpainting trying to approximate the minimum of the proposed energy functional. Our experiments show that the combination of all three terms of the proposed energy works better than taking each term separately, and the results obtained are within the state-of-the-art.