The idea of gradually increasing the capacity of a neural network, during or after training, has gained attention in recent years.Online network growing raises three fundamental questions: when, where, and how to expand.Among all questions, the present paper focuses on the where.We introduce the Normalized Activation Gradient Norm (NAGN), a lightweight criterion to detect under-expressive layers using standard backpropagation signals.Experiments on image classification demonstrate that this approach consistently discovers compact architectures that match larger static baselines with reduced training costs.

International audience

To reduce the computational and memory cost of ConvNets, structured pruning consists in removing features in convolutional and/or fully-connected layers of a pretrained network. As a novel contribution, we study the selection of features as a convex quadratic program with linear constraints. In addition, the updating of weights is treated as the resolution of a linear least-squares system. The quadratic objective function arises from the least-squares reconstruction error on the outputs of pruned layers. Experiments show that our method gives accuracies on par with other reconstruction error-based methods, while achieving a clear gain on computation time.

We propose to tackle dynamic texture video classification as a pattern mining problem. In a nutshell, videos are represented by frequent sequences of representative patches. Firstly, we use a Gaussian Mixture Model to make the clustering of patches from training videos. Secondly, a soft assignment is used as an encoding method to construct sequences of probability vectors (p-sequences) representing sequences of spatio-temporal patches. Thirdly, for each class, we mine meaningful motifs appearing inside the training p-sequences by means of an adapted data mining approach. Finally, feature vectors are constructed from the mined motifs, using the probabilistic support, which quantifies the match between the p-sequences, of the video to be classified, and the key-motifs of the classes. Experimental results and analysis for dynamic texture classification on benchmark datasets (i.e. UCLA, Traffic) show the interest of the proposed method.

We propose a method for human activity recognition from RGB data that does not rely on any pose information during test time, and does not explicitly calculate pose information internally. Instead, a visual attention module learns to predict glimpse sequences in each frame. These glimpses correspond to interest points in the scene that are relevant to the classified activities. No spatial coherence is forced on the glimpse locations, which gives the attention module liberty to explore different points at each frame and better optimize the process of scrutinizing visual information. Tracking and sequentially integrating this kind of unstructured data is a challenge, which we address by separating the set of glimpses from a set of recurrent tracking/recognition workers. These workers receive glimpses, jointly performing subsequent motion tracking and activity prediction. The glimpses are soft-assigned to the workers, optimizing coherence of the assignments in space, time and feature space using an external memory module. No hard decisions are taken, i.e. each glimpse point is assigned to all existing workers, albeit with different importance. Our methods outperform the state-of-the-art on the largest human activity recognition dataset available to-date, NTU RGB+D, and on the Northwestern-UCLA Multiview Action 3D Dataset.

The skeleton is an efficient and complete shape descriptor often used for matching. However, existing skeleton-based shape matching methods are computationally intensive. To reduce the algorithmic complexity, we propose to exploit the natural hierarchy of the skeleton. The aim is to quantify the importance of skeleton branches to guide the shape matching algorithm, in order to match branches having the same order of importance. Our method is based on successive shape smoothing operations and on the deformability of the skeleton to adapt it to each smoothed shape. Moreover, we show that our method is independent from the initial skeleton.

Active contours and minimal paths have been extensively studied theoretical tools for image segmentation. The recent geodesically linked active contour model, which basically consists in a set of vertices connected by paths of minimal cost, blend the bene ts of both concepts. This makes up a closed piecewise-smooth curve, over which an edge or region energy functional can be formulated. As an important shortcoming, the geodesically linked active contour model in its initial formulation does not guarantee the curve to be simple, consistent with respect to the purpose of segmentation. In this paper, we propose to extract a relevant contour from a set of possible paths, such that the resulting structure ts the image data and is simple. Toward this goal, we introduce a novel term to favor the simplicity of the generated contour, as well as a local search method to choose the best combination among possible paths.

In this paper, we propose a specific method for the identification of compound-leaved tree species, with the aim of integrating it in an educational smartphone application. Our work is based on dedicated shape models for compound leaves, designed to estimate the number and shape of leaflets. A deformable template approach is used to fit these models and produce a high-level interpretation of the image content. The resulting models are later used for the segmentation of leaves in both plain and natural background images, by the use of multiple region-based active contours. Combined with other botany-inspired descriptors accounting for the morphological properties of the leaves, we propose a classification method that makes a semantic interpretation possible. Results are presented over a set of more than 1000 images from 17 European tree species, and an integration in the existing mobile application Folia1is considered.

International audience

We address the problem of object tracking within image sequences through region-based energy minimization. A common underlying assumption in region tracking is that color statistics can be confidently estimated in a global manner over object and background regions. This can be a drawback for tracking in real scenes with cluttered backgrounds, where statistical color data is highly scattered, preventing the estimation of reliable color statistics for object/background discrimination. To overcome this limitation, we propose an approach based on a narrow perception of background, which concentrates on the vicinity of tracked objects and thus extract more consistent statistical data for region separation. The benefits of our approach are demonstrated using two different statistical color models.

We describe a 3D+T active surface model for segmentation and tracking of the left ventricular endocardium within 3D MRI sequences of the cardiac cycle. In order to perform segmentation and tracking simultaneously, the surface structure is divided through both time and space, and is therefore handled as an array of planar active contours, interconnected between adjacent slices and frames, providing spatial and temporal consistency. In a given frame, the stacking of slice contours constitute a 3D triangular mesh with a cylindrical topology. Extraction of ventricle border is performed by means of energy minimization, using a combination of boundary-based term and a new computationally efficient region-based term.

We introduce a narrow band region approach in explicit de-formable models for 2D and 3D image segmentation. Embedding a region term into the evolution process, we derive a general formulation which is applied both on a 2D parametric contour and a 3D triangular mesh. Evolution of deformable models is performed by means of energy minimization using the computationally efficient greedy algorithm. The use of a region energy related to the vicinity of the evolving surface overcomes limitations of edge-based active models while remaining time effective. Experiments with segmentation quality assessment are carried out on medical images.

Deformable models, such as the discrete active contour and surface, imply the use of iterative evolution methods to perform 2D and 3D image segmentation. Among the several existing evolution methods, we focus on the greedy algorithm, which minimizes an energy functional, and the physics-based method, which applies forces in order to solve a dynamic differential equation. In this paper, we compare the greedy and physics-based approaches applied on 2D and 3D models, as regards overall speed and segmentation quality, quantified with an evaluating function mainly based on the mean distance between the model and the desired shape.