Romaric Audigier

Morphological Image Processing

Morphological processing (MP) has applications in diverse areas of image processing as filtering, segmentation, and pattern recognition, to both binary and grayscale images. One of the most important operations in morphological image processing is reconstruction from markers. The basic idea is to mark certain image components and then to reconstruct that portion of the image that contains the marked components. The basic fitting operation of morphology is the erosion of an image by a structuring element. Erosion is done by scanning the image with the structuring element. When the structuring element fits completely inside the object, the probe position is marked. The erosion result consists of all scanning locations, where the structuring element fits inside the object. The eroded image is usually a shrunken version of the image, and the shrinking effect is controlled by the structuring element size and shape. As an extension of the binary case, grayscale opening (closing) can be achieved simply by threshold decomposition, followed by binary opening (closing) and stack reconstruction. Grayscale opening and closing have the same properties as their binary equivalents.

Seed-Relative Segmentation Robustness of Watershed and Fuzzy Connectedness Approaches

We define a new class of curves, called geodesic Bezier curves, that are suitable for modeling on manifold triangulations. As a natural generalization of Bezier curves, the new curves are as smooth as possible. We discuss the construction of C0 and C1 piecewise Bezier splines. We also describe how to perform editing operations, such as trimming, using these curves. Special care is taken to achieve interactive rates for modeling tasks.This paper analyzes the robustness issue in three segmentation approaches: the iterative relative fuzzy object extraction, the watershed transforms (WT) by image foresting transform and by minimum spanning forest. These methods need input seeds, which can be source of variability in the segmentation result. So, the robustness of these segmentation methods in relation to the input seeds is focused. The core of each seed is defined as the region where the seed can be moved without altering the segmentation result. We demonstrate that the core is identical for the three methods providing that the tie-zone transform has previously been applied on these methods. Indeed, as the two WT approaches do not return unique solution, the set of possible solutions has to be considered in a unified solution. So does the tie-zone transform. As opposed to what we could think, we show that the core is included in but different from the catchment basin. We also demonstrate that the tie-zone transforms of these WTs are always identical. Furthermore, the framework of minimal sets of seeds, an inverse problem of segmentation, is extended to the pixel level and related to the cores. A new algorithm for the computation of minimal seed sets is finally proposed.

Improving Multi-frame Data Association with Sparse Representations for Robust Near-online Multi-object Tracking

Multiple Object Tracking still remains a difficult problem due to appearance variations and occlusions of the targets or detection failures. Using sophisticated appearance models or performing data association over multiple frames are two common approaches that lead to gain in performances. Inspired by the success of sparse representations in Single Object Tracking, we propose to formulate the multi-frame data association step as an energy minimization problem, designing an energy that efficiently exploits sparse representations of all detections. Furthermore, we propose to use a structured sparsity-inducing norm to compute representations more suited to the tracking context. We perform extensive experiments to demonstrate the effectiveness of the proposed formulation , and evaluate our approach on two public authoritative benchmarks in order to compare it with several state-of-the-art methods.

Duality between the Watershed by Image Foresting Transform and the Fuzzy Connectedness Segmentation Approaches

This paper makes a rereading of two successful image segmentation approaches, the fuzzy connectedness (FC) and the watershed (WS) approaches, by analyzing both by means of the image foresting transform (IFT). This graph-based transform provides a sound framework for analyzing and implementing these methods. This paradigm allows to show the duality existing between the WS by IFT and the FC segmentation approaches. Both can be modeled by an optimal forest computation in a dual form (maximization of the similarities or minimization of the dissimilarities), the main difference being the input parameters: the weights associated to each arc of the graph representing the image. In the WS approach, such weights are based on the (possibly filtered) image gradient values whereas they are based on much more complex affinity values in the FC theory. An efficient algorithm for both FC and IFT-WS computation is proposed. Segmentation robustness issue is also discussed

ONLINE MULTI-PERSON TRACKING BASED ON GLOBAL SPARSE COLLABORATIVE REPRESENTATIONS

Multi-person tracking is still a challenging problem due to recurrent occlusion, pose variation and similar appearances between people. Inspired by the success of sparse representations in single object tracking and face recognition, we propose in this paper an online tracking by detection framework based on collaborative sparse representations. We argue that collaborative representations can better differentiate people compared to target-specific models and therefore help to produce a more robust tracking system. We also show that despite the size of the dictionaries involved, these representations can be efficiently computed with large-scale optimization techniques to get a near real-time algorithm. Experiments show that the proposed approach compares well to other recent online tracking systems on various datasets.

Uniquely-Determined Thinning of the Tie-Zone Watershed Based on Label Frequency

There are many watershed transform algorithms in literature but most of them do not exactly correspond to their respective definition. The solution given by such algorithms depends on their implementation. Others fit with their definition which allows multiple solutions. The solution chosen by such algorithms depends on their implementation too. It is the case of the watershed by image foresting transform that consists in building a forest of trees with minimum path-costs. The recently introduced tie-zone watershed (TZW) has the advantage of not depending on arbitrary implementation choices: it provides a unique and, thereby, unbiased solution. Indeed, the TZW considers all possible solutions of the watershed transform and keeps only the common parts of them as catchment basins whereas parts that differ form a tie zone disputed by many solutions. Although the TZW insures the uniqueness of the solution, it does not prevent from possible large tie zones, sometimes unwanted in segmentation applications. This paper presents a special thinning of the tie zone that leads to a unique solution. Observing all the possible solutions of the watershed by image foresting transform, one can deduce the frequency of the labels associated with each pixel. The thinning consists in assigning the most frequent label while preserving the segmented region connectivity. We demonstrate that the label frequency can be computed both from an immersion-like recursive formula and the proposed fragmented drop paradigm. Finally, we propose an algorithm under the IFT framework that computes the TZW, the label frequency and the thinning simultaneously and without explicit calculation of all the watershed solutions.

3D visualization to assist iterative object definition from medical images

In medical imaging, many applications require visualization and/or analysis of three-dimensional (3D) objects (e.g. organs). At same time, object definition often requires considerable user assistance. In this process, objects are usually defined in an iterative way and their visualization during the process is very important to guide the users actions for the next iteration. The usual procedure provides slice visualization during object definition (segmentation) and 3D visualization afterward. In this paper, we propose and evaluate efficient methods to provide 3D visualization during iterative object definition. The methods combine the differential image foresting transform for segmentation with voxel splatting/ray casting for visualization.

RIMOC, a feature to discriminate unstructured motions

A novel and compact feature discriminating structuredness of observed motions.A feature embedded in a weakly supervised learning framework.An efficient method for real-time violence detection in on-board video-surveillance.Ability of the learned model to generalize training data for varied contexts.A new dataset representative of the targeted application for extensive evaluation. In video-surveillance, violent event detection is of utmost interest. Although action recognition has been well studied in computer vision, literature for violence detection in video is far sparser, and even more for surveillance applications. As aggressive events are difficult to define due to their variability and often need high-level interpretation, we decided to first try to characterize what is frequently present in video with violent human behaviors, at a low level: jerky and unstructured motion. Thus, a novel problem-specific Rotation-Invariant feature modeling MOtion Coherence (RIMOC) was proposed, in order to capture its structure and discriminate the unstructured motions. It is based on the eigenvalues obtained from the second-order statistics of the Histograms of Optical Flow vectors from consecutive temporal instants, locally and densely computed, and further embedded into a spheric Riemannian manifold.The proposed RIMOC feature is used to learn statistical models of normal coherent motions in a weakly supervised manner. A multi-scale scheme applied on an inference-based method allows the events with erratic motion to be detected in space and time, as good candidates of aggressive events.We experimentally show that the proposed method produces results comparable to a state-of-the-art supervised approach, with added simplicity in training and computation. Thanks to the compactness of the feature, real-time computation is achieved in learning as well as in detection phase. Extensive experimental tests on more than 18?h of video are provided in different in-lab and real contexts, such as railway cars equipped with on-board cameras.