Personal space is the physical distance individuals prefer to maintain from others, and its violation induces stress, anxiety, and behavioral change—a phenomenon intensified during the COVID-19 pandemic. While prior research has largely relied on postexperiment questionnaires in passive, one-to-one scenarios, few studies have examined dynamic, crowd-based settings, and to our knowledge, none have collected time-resolved subjective anxiety aligned with moment-to-moment environmental stimuli. We introduce a novel social VR simulation that enables such dynamic analysis by combining continuous self-reports of anxiety with highresolution behavioral tracking during a two-stage, goal-oriented task. In Stage 1, participants navigated immersive VR environments (indoor/outdoor) while encountering virtual agents—some invading personal space, coughing, or (not) wearing masks. In Stage 2, they retrospectively annotated their experiences using first-person video playback. We also administered personality and COVID-related questionnaires. Across$N=109$participants, we collected over 40,000 seconds of annotated data ($\approx 374.56$s per person). Key findings include: 1) Personal space invasions significantly increased anxiety and movement, 2) Coughing agents elevated stress responses, while masks had no significant effect, and 3) Goal achievement consistently reduced anxiety, overriding other stressors. These results provide fine-grained insight into the temporal dynamics of emotional and behavioral responses to social stressors, and offer new design implications for social VR environments.

This study explores how human and AI visual attention differ in a short-term prediction task, particularly in the moments before an accident is about to happen. Since real-world studies of this kind would pose ethical and safety risks, we employed virtual reality (VR) to simulate an accident scenario. In the scenario, the driver approaches a fork in the road, knowing that one path would lead off a cliff crashing the car fatally-as the fork comes closer, the other, safe, path is suddenly blocked by trees, forcing the driver to make a split-second decision where to go. A total of $N=71$ drivers completed the task, and we asked another $N=30$ observers to watch short video clips leading up to the final event and to predict which way the driver would take. We then compared both prediction accuracy as well as attention patterns-how focus is distributed across objects-with AI systems, including vision language models (VLMs) and vision-only models. We found that overall, prediction performance increased as the accident time point approached; interestingly, humans fared better than AI systems overall except for the final time period just before the event. We also found that humans adapted their attention dynamically, shifting focus to important scene elements before an event, whereas AI attention remained static, overlooking key details of the scene. Importantly, as the accident time point approached, human-AI attentional alignment decreased, even though both types of models improved in prediction accuracy. Despite distinct temporal trajectories-vision-only models declining from an early advantage and VLMs peaking in the middle-both models achieved low to zero alignment with human attention. These findings highlight a critical dissociation: AI models make accurate predictions, but rely on visual strategies diverging from human processing, underscoring a gap between explainability and task performance.

What happens if we put vision and touch into conflict? Which modality "wins"? Although several previous studies have addressed this topic, they have solely focused on integration of vision and touch for low-level object properties (such as curvature, slant, or depth). In the present study, we introduce a multimodal mixed-reality setup based on real-time hand-tracking, which was used to display real-world, haptic exploration of objects in a virtual environment through a head-mounted-display (HMD). With this setup we studied multimodal conflict situations of objects varying along higher-level, parametrically-controlled global shape properties. Participants explored these objects in both unimodal and multimodal settings with the latter including congruent and incongruent conditions and differing instructions for weighting the input modalities. Results demonstrated a surprisingly clear touch dominance throughout all experiments, which in addition was only marginally influenceable through instructions to bias their modality weighting. We also present an initial analysis of the hand-tracking patterns that illustrates the potential for our setup to investigate exploration behavior in more detail.

Our work for the first time evaluates the effectiveness of visual and acoustic warning systems in an accident situation using a realistic, immersive driving simulation. In a first experiment, 70 participants were trained to complete a course at high speed. The course contained several forks where a wrong turn would lead to the car falling off a cliff and crashing – these forks were indicated either with a visual warning sign for a first, no-sound group or with a visual and auditory warning cue for a second, sound group. In a testing phase, right after the warning signals were given, trees suddenly fell on the road, leaving the (fatal) turn open. Importantly, in the no-sound group, 18 out of 35 people still chose this turn, whereas in the sound group only 5 out of 35 people did so – the added sound therefore had a large and significant increase in situational awareness. We found no other differences between the groups concerning age, physiological responses, or driving experience. In a second replication experiment, the setup was repeated with another 70 participants without emphasis on driving speed. Results fully confirmed the previous findings with 17 out of 35 people in the no-sound group versus only 6 out of 35 in the sound group choosing the turn to the cliff. With these two experiments using a one-shot design to avoid pre-meditation and testing naïve, rapid decision-making, we provide clear evidence for the advantage of visual-auditory in-vehicle warning systems for promoting situational awareness.

Current supervised methods for facial landmark detection require a large amount of training data and may suffer from overfitting to specific datasets due to the massive number of parameters. We introduce a semi-supervised method in which the crucial idea is to first generate implicit face knowledge from the large amounts of unlabeled images of faces available today. In a first, completely unsupervised stage, we train an adversarial autoencoder to reconstruct faces via a low-dimensional face embedding. In a second, supervised stage, we interleave the decoder with transfer layers to retask the generation of color images to the prediction of landmark heatmaps. Our framework (3FabRec) achieves state-of-the-art performance on several common benchmarks and, most importantly, is able to maintain impressive accuracy on extremely small training sets down to as few as 10 images. As the interleaved layers only add a low amount of parameters to the decoder, inference runs at several hundred FPS on a GPU.

Emergency situations during car driving sometimes force the driver to make a sudden decision. Predicting these decisions will have important applications in updating risk analyses in insurance applications, but also can give insights for drafting autonomous vehicle guidelines. Studying such behavior in experimental settings, however, is limited by ethical issues as it would endanger peoples' lives. Here, we employed the potential of virtual reality (VR) to investigate decision-making in an extreme situation in which participants would have to sacrifice others in order to save themselves. In a VR driving simulation, participants first trained to complete a difficult course with multiple crossroads in which the wrong turn would lead the car to fall down a cliff. In the testing phase, obstacles suddenly appeared on the "safe" turn of a crossroad: for the control group, obstacles consisted of trees, whereas for the experimental group, they were pedestrians. In both groups, drivers had to decide between falling down the cliff or colliding with the obstacles. Results showed that differences in personality traits were able to predict this decision: in the experimental group, drivers who collided with the pedestrians had significantly higher psychopathy and impulsivity traits, whereas impulsivity alone was to some degree predictive in the control group. Other factors like heart rate differences, gender, video game expertise, and driving experience were not predictive of the emergency decision in either group. Our results show that self-interest related personality traits affect decision-making when choosing between preservation of self or others in extreme situations and showcase the potential of virtual reality in studying and modeling human decision-making.

Understanding how humans make decisions in challenging situations - such as trying to save peoples' lives even though this endangers one's own life - is crucial in optimizing rescue operations and dealing with natural disasters and other crises. The experimental study of these decisions, however, has often been done using text-based surveys, which is known to emphasize rational and reflective judgments. Here, we used virtual reality to investigate decision-making in a real-world context, in which a decision needs to be made intuitively under time pressure - for this we simulated an accident situation in an immersive virtual driving scenario. In the scenario, participants were told to race a course as fast as possible. After training, participants were confronted with the sudden appearance of pedestrians on the race course. We observed three different behaviors: group one ignored the pedestrians and/or hit the accelerator, group two hit the brake, and group three tried to steer the car to avoid pedestrians. We found that most Avoid-group participants had more real and virtual driving experience compared to the other two groups and they also felt more competent during the game as measured by subjective game experience questionnaires. Importantly, results from established personality questionnaires showed that participants who did not brake (therefore potentially harming the pedestrians) had significantly lower scores in perspective-taking and higher scores in psychopathy compared to participants who tried to avoid the accident situation. Our results demonstrate that personality differences to some degree are able to predict intuitive decision-making and that such processes can be studied in a controlled, immersive VR simulation.

ABSTRACT We can learn a lot about someone by watching their facial expressions and body language. Harnessing these aspects of non‐verbal communication can lend artificial communication agents greater depth and realism but requires a sound understanding of the relationship between cognition and expressive behaviour. Here, we extend traditional word‐based methodology to use actual videos and then extract the semantic/cognitive space of facial expressions. We find that depending on the specific expressions used, either a four‐dimensional or a two‐dimensional space is needed to describe the variance in the stimuli. The shape and structure of the 4D and 2D spaces are related to each other and very stable to methodological changes. The results show that there is considerable variance between how different people express the same emotion. The recovered space can well capture the full range of facial communication and is very suitable for semantic‐driven facial animation. Copyright © 2014 John Wiley & Sons, Ltd.

For any robot, the ability to recognize and manipulate unknown objects is crucial to successfully work in natural environments. Object recognition and categorization is a very challenging problem, as 3-D objects often give rise to ambiguous, 2-D views. Here, we present a perception-driven exploration and recognition scheme for in-hand object recognition implemented on the iCub humanoid robot. In this setup, the robot actively seeks out object views to optimize the exploration sequence. This is achieved by regarding the object recognition problem as a localization problem. We search for the most likely viewpoint position on the viewsphere of all objects. This problem can be solved efficiently using a particle filter that fuses visual cues with associated motor actions. Based on the state of the filter, we can predict the next best viewpoint after each recognition step by searching for the action that leads to the highest expected information gain. We conduct extensive evaluations of the proposed system in simulation as well as on the actual robot and show the benefit of perception-driven exploration over passive, vision-only processes at discriminating between highly similar objects. We demonstrate that objects are recognized faster and at the same time with a higher accuracy.

Natural image statistics is an important area of research in cognitive sciences and computer vision. Visualization of statistical results can help identify clusters and anomalies as well as analyze deviation, distribution, and correlation. Furthermore, they can provide visual abstractions and symbolism for categorized data. In this paper, we begin our study of visualization of image statistics by considering visual representations of power spectra, which are commonly used to visualize different categories of images. We show that they convey a limited amount of statistical information about image categories and their support for analytical tasks is ineffective. We then introduce several new visual representations, which convey different or more information about image statistics. We apply ANOVA to the image statistics to help select statistically more meaningful measurements in our design process. A task-based user evaluation was carried out to compare the new visual representations with the conventional power spectra plots. Based on the results of the evaluation, we made further improvement of visualizations by introducing composite visual representations of image statistics.

Interaction with its environment is a key requisite for a humanoid robot. Especially the ability to recognize and manipulate unknown objects is crucial to successfully work in natural environments. Visual object recognition, however, still remains a challenging problem, as three-dimensional objects often give rise to ambiguous, two-dimensional views. Here, we propose a perception-driven, multisensory exploration and recognition scheme to actively resolve ambiguities that emerge at certain viewpoints. We define an efficient method to acquire two-dimensional views in an object-centered task space and sample characteristic views on a view sphere. Information is accumulated during the recognition process and used to select actions expected to be most beneficial in discriminating similar objects. Besides visual information we take into account proprioceptive information to create more reliable hypotheses. Simulation and real-world results clearly demonstrate the efficiency of active, multisensory exploration over passive, visiononly recognition methods.

Performance of automatic face recognition algorithm has increased considerably over the past decades. However, face recognition under changes in lighting conditions remains a challenging issue for computers. In this paper, we propose a novel face recognition algorithm inspired by information taken from human fixation patterns. We augment a LGBP (Local Gabor Binary Pattern) algorithm - a well-known face recognition algorithm - to allocate different weights to each facial part during processing. For deriving the weights, we analyzed data from a human face recognition experiment using eye-tracking. Eye-tracking allows us to determine the facial parts during the recognition process which represent salient regions for human processing. Face images are pre-processed during the recognition step using a weight mask based on the salient regions from the eye-tracking data. A comparison with the standard non-weighted LGBP approach demonstrates the efficacy of our method with the weighted method performing better under lighting changes.

In human body pose estimation, manifold learning is a useful method for reducing the dimension of 2D images and 3D body configuration data. Most commonly, body pose is estimated from silhouettes derived from images or image sequences. A major problem when applying manifold estimation, however, is its vulnerability to silhouette variation. In this paper, we propose a novel approach to solving viewpoint-induced silhouette variation by introducing biased label distances for learning manifolds that are able to represent variations in viewpoint, pose, and 3D body configuration. We demonstrate the effectiveness of the approach on a synthetic and a real-world dataset.

A 3D morphable face model represents the complex shape and appearance structure of faces as a compact representation in a vector space. Such a model has proven extremely useful in various fields such as face recognition, facial expression synthesis, face animation, as well as in the perceptual and cognitive sciences. One of the most important steps in creating a 3D morphable model is to establish correspondence between all faces - a highly non-trivial process because of the complex topology of faces. In this paper, we propose to use implicit surfaces for this task, which enables a robust method for establishing dense 3D correspondences. This algorithm is used to build a morphable face model from a large database of Korean faces that were collected. We show that our model is in correspondence with the well-known MPI model so that our face space can be expanded from Asian to Caucasian faces. In addition, with the resulting control parameters of the model, we demonstrate realistic changes of high-level facial attributes.

A good data corpus lies at the heart of progress in both perceptual/cognitive science and in computer vision. While there are a few datasets that deal with simple actions, creating a realistic corpus for complex, long action sequences that contains also human-human interactions has so far not been attempted to our knowledge. Here, we introduce such a corpus for (inter)action understanding that contains six everyday scenarios taking place in a kitchen / living-room setting. Each scenario was acted out several times by different pairs of actors and contains simple object interactions as well as spoken dialogue. In addition, each scenario was first recorded with several HD cameras and also with motion-capturing of the actors and several key objects. Having access to the motion capture data allows not only for kinematic analyses, but also allows for the production of realistic animations where all aspects of the scenario can be fully controlled. We also present results from a first series of perceptual experiments that show how humans are able to infer scenario classes, as well as individual actions and objects from computer animations of everyday situations. These results can serve as a benchmark for future computational approaches that begin to take on complex action understanding.

In this study we want to address the question to what extent the visual and the haptic modalities contribute to the final formation of a complex multisensory perceptual space. By varying three shape parameters a physical shape space of shell-like objects was generated. Participants were allowed to either see or touch the objects or use both senses to explore the objects. Similarity ratings were performed and analyzed using multidimensional scaling (MDS) techniques. By comparing the unimodal perceptual spaces to the multimodal perceptual space we tried to resolve the impact of the visual and the haptic modalities on the combined percept. We found that neither the visual nor the haptic modality dominated the final percept, but rather that the two modalities contributed to the combined percept almost equally. To investigate to which degree these results are transferrable to natural objects, we performed the same visual, haptic, and visuo-haptic similarity ratings and multidimensional scaling analyses using a set of natural sea shells. Again, we found almost equal contributions of the visual and the haptic modalities to the combined percept. Our results suggest that multisensory perceptual spaces are based on a complex combination of object information gathered by different senses.

We present a method for image interpolation that is able to create high-quality, perceptually convincing transitions between recorded images. By implementing concepts derived from human vision, the problem of a physically correct image interpolation is relaxed to that of image interpolation which is perceived as visually correct by human observers. We find that it suffices to focus on exact edge correspondences, homogeneous regions and coherent motion to compute convincing results. A user study confirms the visual quality of the proposed image interpolation approach. We show how each aspect of our approach increases perceived quality of the result. We compare the results to other methods and assess achievable quality for different types of scenes.

In human body pose estimation, manifold learning is a popular technique for reducing the dimension of 2D images and 3D body configuration data. This technique, however, is especially vulnerable to silhouette variation such as caused by viewpoint changes. In this paper, we propose a novel approach that combines three separate manifolds for representing variations in viewpoint, pose and 3D body configuration. We use biased manifold learning to learn these manifolds with appropriately weighted distances. A set of four mapping functions are then learned by a generalized regression neural network for added robustness. Despite using only three manifolds, we show that this method can reliably estimate 3D body poses from 2D images with all learned viewpoints.

Previous work in real and virtual settings has shown that the way in which we interact with objects plays a fundamental role in the way we perceive them. This article uses multidimensional scaling (MDS) analysis to further characterize and quantify the effects of using different haptic exploratory procedures (EPs) on perceptual similarity spaces. In Experiment 1, 20 participants rated similarity on a set of nine novel, 3D objects varying in shape and texture after either following their contours, laterally rubbing their centers, gripping them, or sequentially touching their tips. MDS analysis was used to recover perceptual maps of the objects and relative weights of perceptual dimensions from similarity data. Both the maps and relative weights of shape/texture properties were found to vary as a function of the EP used. In addition, large individual differences in the relative weight of shape/texture were observed. In Experiment 2, 17 of the previous participants repeated Experiment 1 after an average of 105 days. The same patterns of raw similarity ratings, perceptual maps, dimension weights, and individual differences were observed, indicating that perceptual similarities remained stable over time. The findings underscore the role of hand movements and individual biases in shaping haptic perceptual similarity. A framework for validating multimodal virtual displays based on the approach used in the study is also presented.

The human face is capable of producing an astonishing variety of expressions—expressions for which sometimes the smallest difference changes the perceived meaning considerably. Producing realistic-looking facial animations that are able to transmit this degree of complexity continues to be a challenging research topic in computer graphics. One important question that remains to be answered is: When are facial animations good enough? Here we present an integrated framework in which psychophysical experiments are used in a first step to systematically evaluate theperceptualquality of several different computer-generated animations with respect to real-world video sequences. The first experiment provides an evaluation of several animation techniques, exposing specific animation parameters that are important to achieve perceptual fidelity. In a second experiment, we then use these benchmarked animation techniques in the context of perceptual research in order to systematically investigate the spatiotemporal characteristics of expressions. A third and final experiment uses the quality measures that were developed in the first two experiments to examine the perceptual impact of changing facial features to improve the animation techniques. Using such an integrated approach, we are able to provide important insights into facial expressions for both the perceptual and computer graphics community.

Categorization of scenes is a fundamental process of human vision that allows us to efficiently and rapidly analyze our surroundings. Several studies have explored the processes underlying human scene categorization, but they have focused on processing global image information. In this study, we present both psychophysical and computational experiments that investigate the role of local versus global image information in scene categorization. In a first set of human experiments, categorization performance is tested when only local or only global image information is present. Our results suggest that humans rely on local, region-based information as much as on global, configural information. In addition, humans seem to integrate both types of information for intact scene categorization. In a set of computational experiments, human performance is compared to two state-of-the-art computer vision approaches that have been shown to be psychophysically plausible and that model either local or global information. In addition to the influence of local versus global information, in a second series of experiments, we investigated the effect of color on the categorization performance of both the human observers and the computational model. Analysis of the human data suggests that color is an additional channel of perceptual information that leads to higher categorization results at the expense of increased reaction times in the intact condition. However, it does not affect reaction times when only local information is present. When color is removed, the employed computational model follows the relative performance decrease of human observers for each scene category and can thus be seen as a perceptually plausible model for human scene categorization based on local image information.

The goal of stylization is to provide an abstracted representation of an image that highlights specific types of visual information. Recent advances in computer graphics techniques have made it possible to render many varieties of stylized imagery efficiently making stylization into a useful technique, not only for artistic, but also for visualization applications. In this paper, we report results from two sets of experiments that aim at characterizing the perceptual impact and effectiveness of three different stylization techniques in the context of dynamic facial expressions. In the first set of experiments, animated facial expressions are stylized using three common techniques (brush, cartoon, and illustrative stylization) and investigated using different experimental measures. Going beyond the usual questionnaire approach, these experiments compare the techniques according to several criteria ranging from subjective preference to task-dependent measures (such as recognizability, intensity) allowing us to compare behavioral and introspective approaches. The second set of experiments use the same stylization techniques on real-world video sequences in order to compare the effect of stylization on natural and artificial stimuli. Our results shed light on how stylization of image contents affects the perception and subjective evaluation of both real and computer-generated facial expressions.

Virtual Reality (VR) is increasingly being used in industry, medicine, entertainment, education, and research. It is generally critical that the VR setups produce behavior that closely resembles real world behavior. One part of any task is the ability to control our posture. Since postural control is well studied in the real world and is known to be strongly influenced by visual information, it is an ideal metric for examining the behavioral fidelity of VR setups. Moreover, VR-based experiments on postural control can provide fundamental new insights into human perception and cognition. Here, we employ the "swinging room paradigm" to validate a specific VR setup. Furthermore, we systematically examined a larger range of room oscillations than previously studied in any single setup. We also introduce several new methods and analyses that were specifically designed to optimize the detection of synchronous swinging between the observer and the virtual room. The results show that the VR setup has a very high behavioral fidelity and that increases in swinging room amplitude continue to produce increases in body sway even at very large room displacements (+/- 80 cm). Finally, the combination of new methods proved to be a very robust, reliable, and sensitive way of measuring body sway.

The perceived similarity between objects may well vary according to the sensory modality/modalities in which they are experienced, an important consideration for the design of multimodal interfaces. In this study, we present a similarity-based method for comparing the perceptual importance of object properties in touch and in vision and show how the method can also be used to validate computational measures of object properties. Using either vision or touch, human subjects judged the similarity between novel, 3D objects which varied parametrically in shape and texture. Similarities were also computed using a set of state-of-the art 2D and 3D computational measures. Two resolutions of 2D and 3D object data were used for these computations in order to test for scale dependencies. Multidimensional scaling (MDS) was then performed on all similarity data, yielding maps of the stimuli in both perceptual and computational spaces, as well as the relative weight of shape and texture dimensions. For this object set, we found that visual subjects accorded more importance to shape than texture, while haptic subjects weighted them roughly evenly. Fit errors between human and computational maps were then calculated to assess each feature's perceptual validity. Shape-biased features provided good overall fits to the human visual data; however, no single feature yielded a good overall fit to the haptic data, in which we observed large individual differences. This work demonstrates how MDS techniques can be used to evaluate computational object features using the criterion of perceptual similarity. It also demonstrates a way of assessing how the perceptual validity of a feature varies as a function of parameters such as the target modality and the resolution of object data. Potential applications of this method for the design of unimodal and multimodal human---machine interfaces are discussed.

Do we judge similarity between two objects to be the same using touch and vision? We investigated this using psychophysical experiments in which subjects rated the similarity between objects presented either visually or haptically. The stimuli were a family of novel, three-dimensional objects whose microgeometry ("texture") and macrogeometry ("shape") were parametrically varied. Multidimensional scaling of the similarity data was used to reconstruct haptic and visual perceptual spaces. For both modalities, a two-dimensional perceptual space was found whose dimensions clearly corresponded to shape and texture. Interestingly, shape dominated in visual space, whereas both shape and texture were important in haptic space. Furthermore, stimuli clusters were observed in this space, suggesting the emergence of category structure based on similarity relationships. The same category boundaries were confirmed in a visual free sorting experiment. This study reveals differences in object processing across modality and demonstrates an approach for analyzing such differences in multisensory visualizations.

Communication plays a central role in everday life. During an average conversation, information is exchanged in a variety of ways, including through facial motion. Here, we employ a custom, model-based image manipulation technique to selectively “freez ” portions of a face in video recordings in order to determine the areas that are sufficient for proper recognition of nine conversational expressions. The results show that most expressions rely primarily on a single facial area to convey meaning with different expressions using different areas. The results also show that the combination of rigid head, eye, eyebrow, and mouth motions is sufficient to produce expressions that are as easy to recognize as the original, unmanipulated recordings. Finally, the results show that the manipulation technique introduced few perceptible artifacts into the altered video sequences. This fusion of psychophysics and computer graphics techniques provides not only fundamental insights into human perception and cognition, but also yields the basis for a systematic description of what needs to move in order to produce realistic, recognizable conversational facial animations.

Abstract Facial expressions can be used to direct the flow of a conversation as well as to improve the clarity of communication. The critical physical differences between expressions can, however, be small and subtle. Clear presentation of facial expressions in applied settings, then, would seem to require a large conversational agent. Given that visual displays are generally limited in size, the usage of a large conversational agent would reduce the amount of space available for the display of other information. Here, we examine the role of image size in the recognition of facial expressions. The results show that conversational facial expressions can be easily recognized at surprisingly small image sizes. Copyright © 2004 John Wiley & Sons, Ltd.

Regardless of whether the humans involved are virtual or real, well-developed conversational skills are a necessity. The synthesis of interface agents that are not only understandable but also believable can be greatly aided by knowledge of which facial motions are perceptually necessary and sufficient for clear and believable conversational facial expressions. Here, we recorded several core conversational expressions (agreement, disagreement, happiness, sadness, thinking, and confusion) from several individuals, and then psychophysically determined the perceptual ambiguity and believability of the expressions. The results show that people can identify these expressions quite well, although there are some systematic patterns of confusion. People were also very confident of their identifications and found the expressions to be rather believable. The specific pattern of confusions and confidence ratings have strong implications for conversational animation. Finally, the present results provide the information necessary to begin a more fine-grained analysis of the core components of these expressions.

Recent developments in computer vision have shown that local features can provide efficient representations suitable for robust object recognition. Support vector machines have been established as powerful learning algorithms with good generalization capabilities. We combine these two approaches and propose a general kernel method for recognition with local features. We show that the proposed kernel satisfies the Mercer condition and that it is, suitable for many established local feature frameworks. Large-scale recognition results are presented on three different databases, which demonstrate that SVMs with the proposed kernel perform better than standard matching techniques on local features. In addition, experiments on noisy and occluded images show that local feature representations significantly outperform global approaches.