Abstract Objects present in paintings help art history specialists interpret and decode artworks. The analysis of large, digitized artistic collections became feasible thanks to modern object detection approaches. Nevertheless, the use of object detection models typically requires fine-tuning for specific tasks. Therefore, art history specialists remain constrained by the categories of objects in existing labeled artistic datasets when using artificial intelligence methods. This limitation can be overcome by using recent models that combine two modalities: vision and text. Vision-language models have made open-vocabulary detection (OVD) possible, allowing detection without restrictions on the applied categories, in contrast to fixed-vocabulary detection. Recent literature lacks a comprehensive review focusing on OVD in artistic images. In this paper we analyze state-of-the-art models for OVD, analyze their transferability to cultural heritage categories and systematically evaluate them on artistic datasets commonly used in literature. The DEArt and IconArt datasets, which are annotated with cultural heritage-specific categories contain paintings from the 11th to the 20th century. While the Watercolor2K dataset, annotated with common object categories consists of watercolor paintings. Based on our analysis, the OWLv2 model achieved the best performance in both object detection and grounding task scenarios on these datasets. Additionally, we discuss existing challenges of open-vocabulary segmentation in artistic images and future tasks.

The high visual quality of modern deepfakes raises significant concerns about the trustworthiness of digital media and makes facial tampering detection more challenging. Although current deep learning-based deepfake detectors achieve excellent results when tested on deepfake images or image sequences generated using known methods, generalization—where a trained model is tasked with detecting deepfakes created with previously unseen manipulation techniques—is still a major challenge. In this paper, we investigate the impact of training spatial and spatio-temporal deep learning network architectures in the image noise residual domain using spatial rich model (SRM) filters on generalization performance. To this end, we conduct a series of tests on the manipulation methods of the FaceForensics++, DeeperForensics-1.0 and Celeb-DF datasets, demonstrating the value of image noise residuals and temporal feature exploitation in tackling the generalization task.

This work proposes a new formulation of the objects modeling combining geometry and appearance. The object local appearance location is referenced with respect to an invariant which is a geometric landmark. The appearance (shape and texture) is a combination of Harris-Laplace descriptor and local binary pattern (LBP), all is described by the invariant local appearance model (ILAM). We applied the model to describe and learn facial appearances and to recognize them. Given the extracted visual traits from a test image, ILAM model is performed to predict the most similar features to the facial appearance, first, by estimating the highest facial probability, then in terms of LBP Histogram-based measure. Finally, by a geometric computing the invariant allows to locate appearance in the image. We evaluate the model by testing it on different images databases. The experiments show that the model results in high accuracy of detection and provides an acceptable tolerance to the appearance variability.

The skyline, defined as the line separating the sky from other objects on the ground, could provide unique and useful information for a variety of applications. This line was used as a key data, especially, for geo-localization and aerial robotic applications. The particular shape and the geometric features of a skyline may be the identity of the landscape itself. The skyline, once well extracted, could show the silhouette of a famous tower, the mountain peaks, or the landscape topography. In this paper, we proposed a geometric description of the extracted skyline from landscapes. Based on some geometric descriptors, we tried to pick up practical measurements for each skyline. The first proposed approach was the straight lines' classification to differentiate between urban and natural landscapes from their horizon line. The second one is the curvature analysis using a Curvature Scale Space descriptor. This descriptor was used to enhance the first one and to distinguish between natural part and buildings in the same skyline. The results obtained from these geometric description tools were very competitive and they will be the inputs for a classification process.

This paper is placed in the context of video traffic analysis. Typically, hundreds of cameras installed in cities produce a very large amount of data, impossible to handle without automatic processing. For helping traffic controllers to take their decisions, it is important to know in real time, the state of the traffic (number and speed of vehicles on each track segment) but also to have temporal statistics of these measures throughout the day, the week or the year. We can obtain such information by detection and tracking of moving objects in videos, which is a widely studied domain. Nevertheless, most of the automatic video analysis systems face many difficulties: occlusions, variation of the apparent size of objects, illumination changes, etc. In these difficult cases, traditional methods provide only partial trajectories of objects. Rather than trying to make these systems more robust for individual tracking, we propose to aggregate the partial data to build global information on the flow of vehicles in the scene. Specifically, we propose a method which, at first, automatically identifies input Eiand output Xjareas in the scene. Secondly, for each pair Ei, Xj, we record the number ni,jof vehicles that enter the scene in Eiand leave it in Xj.

Real time segmentation of scene into objects and background is really important and represents an initial step of object tracking. Starting from the codebook method we propose some modifications which show significant improvements in most of the normal and also difficult conditions. We include parameter of frequency for accessing, deleting, matching and adding codewords in codebook or to move cache codewords into codebook. We also propose an evaluation method in order to objectively compare several segmentation techniques, based on receiver operating characteristic (ROC) analysis and on precision and recall method. We propose to summarize the quality factor of a method by a single value based on a weighted Euclidean distance or on a harmonic mean between two related characteristics.

In video surveillance applications, PTZ cameras can focus and analyze in details specific zones of the scene. In a computer supervised intrusion detection, a single PTZ camera is unable to visualize the entire scene at once. This article proposes an original solution to this problem, by using an additional spherical mirror. Besides the equations needed to control the PTZ camera, this article presents also a self calibration processes of the camera with the mirror.

This article introduces new similarity measures between trajectories, in order to detect uncommon behaviors. These measures are used to find the most common trajectories in a sequence, using an implicit aggregation method. They may be applied to trajectories of objects tracked in real time. Moreover, by combining one or more measures, it is possible to variate the impact of the temporal dimension - velocity along a trajectory. Our experiments show that the measures are able to properly identify rare trajectories in a video, as well as to detect the most frequent ones.

Automatic detection of a falling person in video sequences has interesting applications in video-surveillance and is an important part of future pervasive home monitoring systems. In this paper, we propose a multiview approach to achieve this goal, where motion is modeled using a layered hidden Markov model (LHMM). The posture classification is performed by a fusion unit, merging the decision provided by the independently processing cameras in a fuzzy logic context. In each view, the fall detection is optimized in a given plane by performing a metric image rectification, making it possible to extract simple and robust features, and being convenient for real-time purpose. A theoretical analysis of the chosen descriptor enables us to define the optimal camera placement for detecting people falling in unspecified situations, and we prove that two cameras are sufficient in practice. Regarding event detection, the LHMM offers a principle way for solving the inference problem. Moreover, the hierarchical architecture decouples the motion analysis into different temporal granularity levels, making the algorithm able to detect very sudden changes, and robust to low-level steps errors.

Properly labeling human body parts in video sequences is essential for robust tracking and motion interpretation frameworks. We propose to perform this task by using Graph Matching. The silhouette skeleton is computed and decomposed into a set of segments corresponding to the different limbs. A Graph capturing the topology of the segments is generated and matched against a 3D model of the human skeleton. The limb identification is carried out for each node of the graph, potentially leading to the absence of correspondence. The method captures the minimal information about the skeleton shape. No assumption about the viewpoint, the human pose, the geometry or the appearance of the limbs is done during the matching process, making the approach applicable to every configuration. Some correspondences that might be ambiguous only relying on topology are enforced by tracking each graph node over time. Several results present the efficiency of the labeling, particularly its robustness to limb detection errors that are likely to occur in real situations because of occlusions or low level system failures. Finally the relevance of the labeling in an overall tracking system is described.

Automatic detection of a falling person in video sequences is an important part of future pervasive home monitoring systems. We propose here a robust method to achieve this goal. Motion is modeled by a hierarchical hidden Markov model (HHMM) whose first layer states are related to the orientation of the tracked person. Finding a consistent way for robustly linking the observation vector to the human poses is the heart of our contribution. In that sense, we carefully study the relationship between angles in the 3D world and their projection onto the image plane. After performing an initial image metric rectification, we derive theoretical properties making it possible to bound the error angle introduced by the image formation process for a standing posture. This allows us to confidently identify other poses as "non-standing" ones, and thus to robustly analyze pose sequences against a given motion model. Several results illustrate the efficiency of the algorithm by pointing out its ability to accurately recognize a person falling down from another walking or sitting, as well as its capacity to run in an unspecified configuration

We propose an original method for tracking people based on the construction of a 2-D human appearance model. The general framework, which is a region-based tracking approach, is applicable to any type of object. We show how to specialize the method for taking advantage of the structural properties of the human body. We segment its visible parts, construct and update the appearance model. This latter one provides a discriminative feature capturing both color and shape properties of the different limbs, making it possible to recognize people after they have temporarily disappeared. The method does not make use of skin color detection, which allows us to perform tracking under any viewpoint. The only assumption for the recognition is the approximate viewpoint correspondence during the matching process between the different models. Several results in complex situations prove the efficiency of the algorithm, which runs in near real time. Finally, the model provides an important clue for further human motion analysis process.

The parallel implementation of image processing algorithms implies an important choice of data distribution strategy. In order to handle the specific constraints associated with images, data distribution must take into account not only the locality of the data and its geometrical regularity but also the possible irregular computation costs associated with different image elements. A widely studied field to tackle this problem is the family of methods related to rectilinear partitioning. We introduce two fully parallel heuristics that compute suboptimal partitions, with a better complexity than the best known algorithms that compute optimal partitions. In this paper, we compare our heuristics to an optimal partitioning, both in terms of execution time and accuracy of the partition. We give some theoretical bounds on the quality of these heuristics that are corroborated by results of random numerical experiments and real applications.

This paper introduces and compares three parallel algorithms to compute general geometric image transformations on MIMD machines. We propose three variants of a parallel general scheme. We focus on the load balancing and the data redistributions. Experimental results are reported and compared. The implementation has been done using PPCM library allowing us to run the program over different parallel machines. We compare logical communication schemes for message-passing machines. Since our parallel algorithm needs global communications such as multiscatters, we study the efficiency of two different logical topologies usable with PPCM. These studies allow us to find the best combination of algorithm and virtual topology to use on a given parallel machine.

This paper presents a load-balanced parallelization of the well known Marching-Cubes algorithm, that aims at constructing an iso-surface in a 3D image. We first derive a modelization for the computation time as a function of the generated surface complexity. The workload associated to each slice of the input data is evaluated by counting the number of vertices that will be generated on that slice. The slices are then locally redistributed to ensure a balanced workload. We give an upper bound on the number of polygons of the triangulation, and present a family of surfaces whose number of triangles tends to this bound. This analysis allows us to foresee (and thus to allocate) the memory size needed for the data structures and to assign to each vertex a unique global reference. Experiments done on an Intel Paragon machine are given both for synthetic and medical images. They show the usefulness of our dynamic data redistribution scheme.

This paper discusses several data allocation strategies used for the parallel implementation of basic imaging operators. It shows that depending on the operator (sequential or parallel, with regular or irregular execution time), the image data must be partitioned in very different manners: The square sub-domains are best adapted for minimizing the communication volume, but rectangles can perform better when we take into account the time for constructing messages. Block allocations are well adapted for inherently parallel operators since they minimize interprocessor interactions, but in the case of recursive operators, they lead to nearly sequential executions. In this framework, we show the usefulness of block-cyclic allocations. Finally, we illustrate the fact that allocating the same amount of image data to each processor can lead to severe load imbalance in the case of some operators with data-dependant execution times.

This paper describes the parallel implementation of the Z-Buffer algorithm on a distributed memory machine. The Z-Buffer is one of the most popular techniques used to generate a representation of a scene consisting of objects in a three-dimensional world. We propose and compare two different parallel implementations on a reconfigurable network of Transputers. In the first approach, the description of the scene is distributed among the processors configured as a tree. The picture is processed in a pipelined fashion, in order to output parts of the image during the computation of the remainder. We show the influence of the degree and the height of the tree on the global performance of the algorithm. In a second approach, both the picture and the scene description are distributed to the processors. We have therefore to redistribute dynamically the tiles among the processors at the beginning of the computation. To perform this redistribution, a special algorithm is designed for the case where the processors are configured as a unidirectional or bidirectional ring. Then we implement a greedy algorithm that enables us to perform the redistribution on an arbitrary interconnection network. We show that the two approaches are complementary: for small pictures or large scenes, a tree-based algorithm performs better than a redistribution-based algorithm, but for large pictures or smaller scenes, it is the other way round. We obtain substantial speedups over the sequential implementation, with up to 32 processors.

Performing reduction operations with distributed memory machines whose interconnection networks are reconfigurable is considered. The focus is on machines whose interconnection graph can be configured as any graph of maximum degree d. The best way of interconnecting the p processors as a function of p,d and some problem- and machine-dependent parameters that characterize the ratio communication/arithmetic for the reduction operation are discussed. Experiments on transputer-based networks are in good accordance with the theoretical results.>