Representation Of Visual Scene Research Articles

Enhancing scene understanding based on deep learning for end-to-end autonomous driving

Efficient understanding of the environment is a crucial prerequisite for autonomous driving, but explicitly modeling the environment is hard to come true. In contrast, imitation learning, in theory, can arrive at the direct mapping from visual input to driving command, but the inscrutability of scene representation in imitation learning is still a challenging problem. In this paper, we propose to enhance the abstract representation of visual scene from two aspects for better scene understanding, i.e. Visual Guide path and Driving Affordances path. For Visual Guide path, we leverage semantic information as visual priors to learn the intuitive state of the environment, e.g. the spatial semantic occupation of the visual scene. For Driving Affordances path, several driving affordance indicators reflecting the relationship between environment and vehicle behavior are learned as the global guidance to guide the driving system to learn safe and efficient driving policies. With the complementarity of these two paths, a Bilateral Guide Network is designed to realize the complete mapping from visual input to driving command. Our method is evaluated on the CARLA simulator with various scenarios to demonstrate the effectiveness. Besides, comparative analyses are made with some state-of-the-art methods to justify the performance of our method in the aspect of autonomous driving.

Integrating Multihub Driven Attention Mechanism and Big Data Analytics for Virtual Representation of Visual Scenes

Digital twin is the innovation backbone of the smart manufacturing by delivering virtual representation of the real world. Aiming at constructing virtual representations of visual scenes, scene graph generation is a digital twin task that not only models objects but also infers their relationships. Existing works usually learn coarse global context when predicting relationships leading to excessive redundant information being considered. In this article, we first classify objects into different subgroups according to the degree of correlations with several hub objects. Then, we propose a multihub driven attention network (MHDANet) based on deep learning that drives the information to pass within the subgroups and forces objects to attend more to related objects. Consequently, MHDANet learns compact relation-aware features of visual scenes and predicts accurate and diverse relationships. Experimental results show that MHDANet achieves superb performance on scene graph generation on real-world datasets and especially alleviates the imbalance of predicted relationship categories.

AAP-MIT: Attentive Atrous Pyramid Network and Memory Incorporated Transformer for Multisentence Video Description.

Generating multi-sentence descriptions for video is considered to be the most complex task in computer vision and natural language understanding due to the intricate nature of video-text data. With the recent advances in deep learning approaches, the multi-sentence video description has achieved an impressive progress. However, learning rich temporal context representation of visual sequences and modelling long-term dependencies of natural language descriptions is still a challenging problem. Towards this goal, we propose an Attentive Atrous Pyramid network and Memory Incorporated Transformer (AAP-MIT) for multi-sentence video description. The proposed AAP-MIT incorporates the effective representation of visual scene by distilling the most informative and discriminative spatio-temporal features of video data at multiple granularities and further generates the highly summarized descriptions. Profoundly, we construct AAP-MIT with three major components: i) a temporal pyramid network, which builds the temporal feature hierarchy at multiple scales by convolving the local features at temporal space, ii) a temporal correlation attention to learn the relations among various temporal video segments, and iii) the memory incorporated transformer, which augments the new memory block in language transformer to generate highly descriptive natural language sentences. Finally, the extensive experiments on ActivityNet Captions and YouCookII datasets demonstrate the substantial superiority of AAP-MIT over the existing approaches.

Coding of Navigational Distance and Functional Constraint of Boundaries in the Human Scene-Selective Cortex.

For visually guided navigation, the use of environmental cues is essential. Particularly, detecting local boundaries that impose limits to locomotion and estimating their location is crucial. In a series of three fMRI experiments, we investigated whether there is a neural coding of navigational distance in the human visual cortex (both female and male). We used virtual reality software to systematically manipulate the distance from a viewer perspective to different types of a boundary. Using a multivoxel pattern classification employing a linear support vector machine, we found that the occipital place area (OPA) is sensitive to the navigational distance restricted by the transparent glass wall. Further, the OPA was sensitive to a non-crossable boundary only, suggesting an importance of the functional constraint of a boundary. Together, we propose the OPA as a perceptual source of external environmental features relevant for navigation.SIGNIFICANCE STATEMENT One of major goals in cognitive neuroscience has been to understand the nature of visual scene representation in human ventral visual cortex. An aspect of scene perception that has been overlooked despite its ecological importance is the analysis of space for navigation. One of critical computation necessary for navigation is coding of distance to environmental boundaries that impose limit on navigator's movements. This paper reports the first empirical evidence for coding of navigational distance in the human visual cortex and its striking sensitivity to functional constraint of environmental boundaries. Such finding links the paper to previous neurological and behavioral works that emphasized the distance to boundaries as a crucial geometric property for reorientation behavior of children and other animal species.

Predicting eye movement patterns from fMRI responses to natural scenes

Eye tracking has long been used to measure overt spatial attention, and computational models of spatial attention reliably predict eye movements to natural images. However, researchers lack techniques to noninvasively access spatial representations in the human brain that guide eye movements. Here, we use functional magnetic resonance imaging (fMRI) to predict eye movement patterns from reconstructed spatial representations evoked by natural scenes. First, we reconstruct fixation maps to directly predict eye movement patterns from fMRI activity. Next, we use a model-based decoding pipeline that aligns fMRI activity to deep convolutional neural network activity to reconstruct spatial priority maps and predict eye movements in a zero-shot fashion. We predict human eye movement patterns from fMRI responses to natural scenes, provide evidence that visual representations of scenes and objects map onto neural representations that predict eye movements, and find a novel three-way link between brain activity, deep neural network models, and behavior.

Effect of contextual knowledge on spatial layout extrapolation

Boundary extension (BE) refers to the tendency to remember a previously perceived scene with a greater spatial expanse. This phenomenon is described as resulting from different sources of information: external (i.e., visual) and internally driven (i.e., amodal, conceptual, and contextual) information. Although the literature has emphasized the role of top-down expectations to account for layout extrapolation, their effect has rarely been tested experimentally. In this research, we attempted to determine how visual context affects BE, as a function of scene exposure duration (long, short). To induce knowledge about visual context, the memorization phase of the camera distance paradigm was preceded by a preexposure phase, during which each of the to-be-memorized scenes was presented in a larger spatial framework. In an initial experiment, we examined the effect of contextual knowledge with presentation duration, allowing for in-depth processing of visual information during encoding (i.e., 15 s). The results indicated that participants exposed to the preexposure showed decreased BE, and displayed no directional memory error in some conditions. Because the effect of context is known to occur at an early stage of scene perception, in a second experiment we sought to determine whether the effect of a preview occurs during the first fixation on a visual scene. The results indicated that BE seems not to be modulated by this factor at very brief presentation durations. These results are discussed in light of current visual scene representation theories.

A new approach to solving the feature-binding problem in primate vision.

We discuss a recently proposed approach to solve the classic feature-binding problem in primate vision that uses neural dynamics known to be present within the visual cortex. Broadly, the feature-binding problem in the visual context concerns not only how a hierarchy of features such as edges and objects within a scene are represented, but also the hierarchical relationships between these features at every spatial scale across the visual field. This is necessary for the visual brain to be able to make sense of its visuospatial world. Solving this problem is an important step towards the development of artificial general intelligence. In neural network simulation studies, it has been found that neurons encoding the binding relations between visual features, known as binding neurons, emerge during visual training when key properties of the visual cortex are incorporated into the models. These biological network properties include (i) bottom-up, lateral and top-down synaptic connections, (ii) spiking neuronal dynamics, (iii) spike timing-dependent plasticity, and (iv) a random distribution of axonal transmission delays (of the order of several milliseconds) in the propagation of spikes between neurons. After training the network on a set of visual stimuli, modelling studies have reported observing the gradual emergence of polychronization through successive layers of the network, in which subpopulations of neurons have learned to emit their spikes in regularly repeating spatio-temporal patterns in response to specific visual stimuli. Such a subpopulation of neurons is known as a polychronous neuronal group (PNG). Some neurons embedded within these PNGs receive convergent inputs from neurons representing lower- and higher-level visual features, and thus appear to encode the hierarchical binding relationship between features. Neural activity with this kind of spatio-temporal structure robustly emerges in the higher network layers even when neurons in the input layer represent visual stimuli with spike timings that are randomized according to a Poisson distribution. The resulting hierarchical representation of visual scenes in such models, including the representation of hierarchical binding relations between lower- and higher-level visual features, is consistent with the hierarchical phenomenology or subjective experience of primate vision and is distinct from approaches interested in segmenting a visual scene into a finite set of objects.

Panoramic Memory Shapes Visual Representations of Scenes

Panoramic Memory Shapes Visual Representations of Scenes

The Neural Representation of Multiple Objects in the Primate Visual System.

One of the main tasks of the visual system is to combine the edges and surfaces of individual objects into a perceptual group, and thus create a representation of visual scenes in which multiple objects are segregated from the background. Many studies have focused on how single objects (usually

Decoding individual natural scene representations during perception and imagery

We used a multi-voxel classification analysis of functional magnetic resonance imaging (fMRI) data to determine to what extent item-specific information about complex natural scenes is represented in several category-selective areas of human extrastriate visual cortex during visual perception and visual mental imagery. Participants in the scanner either viewed or were instructed to visualize previously memorized natural scene exemplars, and the neuroimaging data were subsequently subjected to a multi-voxel pattern analysis (MVPA) using a support vector machine (SVM) classifier. We found that item-specific information was represented in multiple scene-selective areas: the occipital place area (OPA), parahippocampal place area (PPA), retrosplenial cortex (RSC), and a scene-selective portion of the precuneus/intraparietal sulcus region (PCu/IPS). Furthermore, item-specific information from perceived scenes was re-instantiated during mental imagery of the same scenes. These results support findings from previous decoding analyses for other types of visual information and/or brain areas during imagery or working memory, and extend them to the case of visual scenes (and scene-selective cortex). Taken together, such findings support models suggesting that reflective mental processes are subserved by the re-instantiation of perceptual information in high-level visual cortex. We also examined activity in the fusiform face area (FFA) and found that it, too, contained significant item-specific scene information during perception, but not during mental imagery. This suggests that although decodable scene-relevant activity occurs in FFA during perception, FFA activity may not be a necessary (or even relevant) component of one's mental representation of visual scenes.

Multi-Scale Spatial Concatenations of Local Features in Natural Scenes and Scene Classification

How does the visual system encode natural scenes? What are the basic structures of natural scenes? In current models of scene perception, there are two broad feature representations, global and local representations. Both representations are useful and have some successes; however, many observations on human scene perception seem to point to an intermediate-level representation.In this paper, we proposed natural scene structures, i.e., multi-scale spatial concatenations of local features, as an intermediate-level representation of natural scenes. To compile the natural scene structures, we first sampled a large number of multi-scale circular scene patches in a hexagonal configuration. We then performed independent component analysis on the patches and classified the independent components into a set of clusters using the K-means method. Finally, we obtained a set of natural scene structures, each of which is characterized by a set of dominant clusters of independent components.We examined a range of statistics of the natural scene structures, compiled from two widely used datasets of natural scenes, and modeled their spatial arrangements at larger spatial scales using adjacency matrices. We found that the natural scene structures include a full range of concatenations of visual features in natural scenes, and can be used to encode spatial information at various scales. We then selected a set of natural scene structures with high information, and used the occurring frequencies and the eigenvalues of the adjacency matrices to classify scenes in the datasets. We found that the performance of this model is comparable to or better than the state-of-the-art models on the two datasets. These results suggest that the natural scene structures are a useful intermediate-level representation of visual scenes for our understanding of natural scene perception.

Template Construction Grammar: From Visual Scene Description to Language Comprehension and Agrammatism

How does the language system coordinate with our visual system to yield flexible integration of linguistic, perceptual, and world-knowledge information when we communicate about the world we perceive? Schema theory is a computational framework that allows the simulation of perceptuo-motor coordination programs on the basis of known brain operating principles such as cooperative computation and distributed processing. We present first its application to a model of language production, SemRep/TCG, which combines a semantic representation of visual scenes (SemRep) with Template Construction Grammar (TCG) as a means to generate verbal descriptions of a scene from its associated SemRep graph. SemRep/TCG combines the neurocomputational framework of schema theory with the representational format of construction grammar in a model linking eye-tracking data to visual scene descriptions. We then offer a conceptual extension of TCG to include language comprehension and address data on the role of both world knowledge and grammatical semantics in the comprehension performances of agrammatic aphasic patients. This extension introduces a distinction between heavy and light semantics. The TCG model of language comprehension offers a computational framework to quantitatively analyze the distributed dynamics of language processes, focusing on the interactions between grammatical, world knowledge, and visual information. In particular, it reveals interesting implications for the understanding of the various patterns of comprehension performances of agrammatic aphasics measured using sentence-picture matching tasks. This new step in the life cycle of the model serves as a basis for exploring the specific challenges that neurolinguistic computational modeling poses to the neuroinformatics community.

Dynamic Visual Representations of Scenes and Objects: The Forest to the Tree

Dynamic Visual Representations of Scenes and Objects: The Forest to the Tree

Exploiting spatial descriptions in visual scene analysis

The reliable automatic visual recognition of indoor scenes with complex object constellations using only sensor data is a nontrivial problem. In order to improve the construction of an accurate semantic 3D model of an indoor scene, we exploit human-produced verbal descriptions of the relative location of pairs of objects. This requires the ability to deal with different spatial reference frames (RF) that humans use interchangeably. In German, both the intrinsic and relative RF are used frequently, which often leads to ambiguities in referential communication. We assume that there are certain regularities that help in specific contexts. In a first experiment, we investigated how speakers of German describe spatial relationships between different pieces of furniture. This gave us important information about the distribution of the RFs used for furniture-predicate combinations, and by implication also about the preferred spatial predicate. The results of this experiment are compiled into a computational model that extracts partial orderings of spatial arrangements between furniture items from verbal descriptions. In the implemented system, the visual scene is initially scanned by a 3D camera system. From the 3D point cloud, we extract point clusters that suggest the presence of certain furniture objects. We then integrate the partial orderings extracted from the verbal utterances incrementally and cumulatively with the estimated probabilities about the identity and location of objects in the scene, and also estimate the probable orientation of the objects. This allows the system to significantly improve both the accuracy and richness of its visual scene representation.

Rethinking visual scene perception.

A classic puzzle in understanding visual scene perception is how to reconcile the physiological constraints of vision with the phenomenology of seeing. Vision captures information via discrete eye fixations, interrupted by saccadic suppression, and limited by retinal inhomogeneity. Yet scenes are effortlessly perceived as coherent, continuous, and meaningful. Two conceptualizations of scene representation will be contrasted. The traditional visual-cognitive model casts visual scene representation as an imperfect reflection of the visual sensory input alone. By contrast, a new multisource model casts visual scene representation in terms of an egocentric spatial framework that is 'filled-in' by visual sensory input, but also by amodal perception, and by expectations and by constraints derived from rapid-scene classification and object-to-context associations. Together, these nonvisual sources serve to 'simulate' a likely surrounding scene that the visual input only partially reveals. Pros and cons of these alternative views will be discussed. WIREs Cogn Sci 2012, 3:117-127. doi: 10.1002/wcs.149 For further resources related to this article, please visit the WIREs website.

Representation of visual scenes by local neuronal populations in layer 2/3 of mouse visual cortex

How are visual scenes encoded in local neural networks of visual cortex? In rodents, visual cortex lacks a columnar organization so that processing of diverse features from a spot in visual space could be performed locally by populations of neighboring neurons. To examine how complex visual scenes are represented by local microcircuits in mouse visual cortex we measured visually evoked responses of layer 2/3 neuronal populations using 3D two-photon calcium imaging. Both natural and artificial movie scenes (10 seconds duration) evoked distributed and sparsely organized responses in local populations of 70–150 neurons within the sampled volumes. About 50% of neurons showed calcium transients during visual scene presentation, of which about half displayed reliable temporal activation patterns. The majority of the reliably responding neurons were activated primarily by one of the four visual scenes applied. Consequently, single-neurons performed poorly in decoding, which visual scene had been presented. In contrast, high levels of decoding performance (>80%) were reached when considering population responses, requiring about 80 randomly picked cells or 20 reliable responders. Furthermore, reliable responding neurons tended to have neighbors sharing the same stimulus preference. Because of this local redundancy, it was beneficial for efficient scene decoding to read out activity from spatially distributed rather than locally clustered neurons. Our results suggest a population code in layer 2/3 of visual cortex, where the visual environment is dynamically represented in the activation of distinct functional sub-networks.

Decoding individual natural scene representations during perception and imagery

We used a multi-voxel classification analysis of functional magnetic resonance imaging (fMRI) data to determine to what extent item-specific information about complex natural scenes is represented in several category-selective areas of human extrastriate visual cortex during visual perception and visual mental imagery. Participants in the scanner either viewed or were instructed to visualize previously memorized natural scene exemplars, and the neuroimaging data were subsequently subjected to a multi-voxel pattern analysis (MVPA) using a support vector machine (SVM) classifier. We found that item-specific information was represented in multiple scene-selective areas: the occipital place area (OPA), parahippocampal place area (PPA), retrosplenial cortex (RSC), and a scene-selective portion of the precuneus/intraparietal sulcus region (PCu/IPS). Furthermore, item-specific information from perceived scenes was re-instantiated during mental imagery of the same scenes. These results support findings from previous decoding analyses for other types of visual information and/or brain areas during imagery or working memory, and extend them to the case of visual scenes (and scene-selective cortex). Taken together, such findings support models suggesting that reflective mental processes are subserved by the re-instantiation of perceptual information in high-level visual cortex. We also examined activity in the fusiform face area (FFA) and found that it, too, contained significant item-specific scene information during perception, but not during mental imagery. This suggests that although decodable scene-relevant activity occurs in FFA during perception, FFA activity may not be a necessary (or even relevant) component of one’s mental representation of visual scenes.

The advantage of a ground surface in the representation of visual scenes

The present study used change detection tasks to examine whether there is an advantage of a ground surface in representing visual scenes. In 6 experiments, a flicker paradigm (Experiments 1 through 4) or a one-shot paradigm (Experiments 5 and 6) was used to examine whether changes on a ground surface were easier to detect than changes on a ceiling surface. Overall, we found that: (1) there was an advantage in detecting changes on a ground surface or changes to objects on a ground surface; (2) this advantage was dependent on the presence of a coherent ground surface; (3) this advantage could propagate to objects connected to the ground surface through "nested" contact relations; (4) this advantage was mainly due to improved encoding rather than improved retrieval and comparison of the ground surface; and (5) this advantage was dependent on the presentation duration of the scene but not the number of objects presented in the scene. Together, these results suggest a unique role of the ground surface in organizing visual scenes.

Representing the Forest before the Trees: A Global Advantage Effect in Monkey Inferotemporal Cortex

Hierarchical stimuli (large shapes composed of small shapes) have long been used to study how humans perceive the global and the local content of a scene--the forest and the trees. Studies using these stimuli have revealed a global advantage effect: humans consistently report global shape faster than local shape. The neuronal underpinnings of this effect remain unclear. Here we demonstrate a correlate and possible mechanism in monkey inferotemporal cortex (IT). Inferotemporal neurons signal the global content of a hierarchical display approximately 30 ms before they signal its local content. This is a specific expression of a general principle, related to spatial scale or spatial frequency rather than to hierarchical level, whereby the representation of a large shape develops in IT before that of a small shape. These findings provide support for a coarse-to-fine model of visual scene representation.

Visual representations of scenes and objects: retinotopical or non-retinotopical?

Visual representations of scenes and objects: retinotopical or non-retinotopical?