Michelle R. Greene

Recognition of natural scenes from global properties: Seeing the forest without representing the trees

Human observers are able to rapidly and accurately categorize natural scenes, but the representation mediating this feat is still unknown. Here we propose a framework of rapid scene categorization that does not segment a scene into objects and instead uses a vocabulary of global, ecological properties that describe spatial and functional aspects of scene space (such as navigability or mean depth). In Experiment 1, we obtained ground truth rankings on global properties for use in Experiments 2-4. To what extent do human observers use global property information when rapidly categorizing natural scenes? In Experiment 2, we found that global property resemblance was a strong predictor of both false alarm rates and reaction times in a rapid scene categorization experiment. To what extent is global property information alone a sufficient predictor of rapid natural scene categorization? In Experiment 3, we found that the performance of a classifier representing only these properties is indistinguishable from human performance in a rapid scene categorization task in terms of both accuracy and false alarms. To what extent is this high predictability unique to a global property representation? In Experiment 4, we compared two models that represent scene object information to human categorization performance and found that these models had lower fidelity at representing the patterns of performance than the global property model. These results provide support for the hypothesis that rapid categorization of natural scenes may not be mediated primarily though objects and parts, but also through global properties of structure and affordance.

The Briefest of Glances The Time Course of Natural Scene Understanding

What information is available from a brief glance at a novel scene? Although previous efforts to answer this question have focused on scene categorization or object detection, real-world scenes contain a wealth of information whose perceptual availability has yet to be explored. We compared image exposure thresholds in several tasks involving basic-level categorization or global-property classification. All thresholds were remarkably short: Observers achieved 75%-correct performance with presentations ranging from 19 to 67 ms, reaching maximum performance at about 100 ms. Global-property categorization was performed with significantly less presentation time than basic-level categorization, which suggests that there exists a time during early visual processing when a scene may be classified as, for example, a large space or navigable, but not yet as a mountain or lake. Comparing the relative availability of visual information reveals bottlenecks in the accumulation of meaning. Understanding these bottlenecks provides critical insight into the computations underlying rapid visual understanding.

Disentangling scene content from spatial boundary: complementary roles for the parahippocampal place area and lateral occipital complex in representing real-world scenes.

Behavioral and computational studies suggest that visual scene analysis rapidly produces a rich description of both the objects and the spatial layout of surfaces in a scene. However, there is still a large gap in our understanding of how the human brain accomplishes these diverse functions of scene understanding. Here we probe the nature of real-world scene representations using multivoxel functional magnetic resonance imaging pattern analysis. We show that natural scenes are analyzed in a distributed and complementary manner by the parahippocampal place area (PPA) and the lateral occipital complex (LOC) in particular, as well as other regions in the ventral stream. Specifically, we study the classification performance of different scene-selective regions using images that vary in spatial boundary and naturalness content. We discover that, whereas both the PPA and LOC can accurately classify scenes, they make different errors: the PPA more often confuses scenes that have the same spatial boundaries, whereas the LOC more often confuses scenes that have the same content. By demonstrating that visual scene analysis recruits distinct and complementary high-level representations, our results testify to distinct neural pathways for representing the spatial boundaries and content of a visual scene.

Reconsidering Yarbus: a failure to predict observers' task from eye movement patterns.

In 1967, Yarbus presented qualitative data from one observer showing that the patterns of eye movements were dramatically affected by an observers task, suggesting that complex mental states could be inferred from scan paths. The strong claim of this very influential finding has never been rigorously tested. Our observers viewed photographs for 10s each. They performed one of four image-based tasks while eye movements were recorded. A pattern classifier, given features from the static scan paths, could identify the image and the observer at above-chance levels. However, it could not predict a viewers task. Shorter and longer (60s) viewing epochs produced similar results. Critically, human judges also failed to identify the tasks performed by the observers based on the static scan paths. The Yarbus finding is evocative, and while it is possible an observers mental state might be decoded from some aspect of eye movements, static scan paths alone do not appear to be adequate to infer complex mental states of an observer.

Statistics of high-level scene context.

Context is critical for recognizing environments and for searching for objects within them: contextual associations have been shown to modulate reaction time and object recognition accuracy, as well as influence the distribution of eye movements and patterns of brain activations. However, we have not yet systematically quantified the relationships between objects and their scene environments. Here I seek to fill this gap by providing descriptive statistics of object-scene relationships. A total of 48, 167 objects were hand-labeled in 3499 scenes using the LabelMe tool (Russell et al., 2008). From these data, I computed a variety of descriptive statistics at three different levels of analysis: the ensemble statistics that describe the density and spatial distribution of unnamed “things” in the scene; the bag of words level where scenes are described by the list of objects contained within them; and the structural level where the spatial distribution and relationships between the objects are measured. The utility of each level of description for scene categorization was assessed through the use of linear classifiers, and the plausibility of each level for modeling human scene categorization is discussed. Of the three levels, ensemble statistics were found to be the most informative (per feature), and also best explained human patterns of categorization errors. Although a bag of words classifier had similar performance to human observers, it had a markedly different pattern of errors. However, certain objects are more useful than others, and ceiling classification performance could be achieved using only the 64 most informative objects. As object location tends not to vary as a function of category, structural information provided little additional information. Additionally, these data provide valuable information on natural scene redundancy that can be exploited for machine vision, and can help the visual cognition community to design experiments guided by statistics rather than intuition.

Visual Scenes are Categorized by Function

How do we know that a kitchen is a kitchen by looking? Traditional models posit that scene categorization is achieved through recognizing necessary and sufficient features and objects, yet there is little consensus about what these may be. However, scene categories should reflect how we use visual information. Therefore, we test the hypothesis that scene categories reflect functions, or the possibilities for actions within a scene. Our approach is to compare human categorization patterns with predictions made by both functions and alternative models. We collected a large-scale scene category distance matrix (5 million trials) by asking observers to simply decide whether 2 images were from the same or different categories. Using the actions from the American Time Use Survey, we mapped actions onto each scene (1.4 million trials). We found a strong relationship between ranked category distance and functional distance (r = .50, or 66% of the maximum possible correlation). The function model outperformed alternative models of object-based distance (r = .33), visual features from a convolutional neural network (r = .39), lexical distance (r = .27), and models of visual features. Using hierarchical linear regression, we found that functions captured 85.5% of overall explained variance, with nearly half of the explained variance captured only by functions, implying that the predictive power of alternative models was because of their shared variance with the function-based model. These results challenge the dominant school of thought that visual features and objects are sufficient for scene categorization, suggesting instead that a scenes category may be determined by the scenes function.

Visual categorization is automatic and obligatory: evidence from Stroop-like paradigm.

Human observers categorize visual stimuli with remarkable efficiency--a result that has led to the suggestion that object and scene categorization may be automatic processes. We tested this hypothesis by presenting observers with a modified Stroop paradigm in which object or scene words were presented over images of objects or scenes. Terms were either congruent or incongruent with the images. Observers classified the words as being object or scene terms while ignoring images. Classifying a word on an incongruent image came at a cost for both objects and scenes. Furthermore, automatic processing was observed for entry-level scene categories, but not superordinate-level categories, suggesting that not all rapid categorizations are automatic. Taken together, we have demonstrated that entry-level visual categorization is an automatic and obligatory process.

What you see is what you expect: rapid scene understanding benefits from prior experience

Although we are able to rapidly understand novel scene images, little is known about the mechanisms that support this ability. Theories of optimal coding assert that prior visual experience can be used to ease the computational burden of visual processing. A consequence of this idea is that more probable visual inputs should be facilitated relative to more unlikely stimuli. In three experiments, we compared the perceptions of highly improbable real-world scenes (e.g., an underwater press conference) with common images matched for visual and semantic features. Although the two groups of images could not be distinguished by their low-level visual features, we found profound deficits related to the improbable images: Observers wrote poorer descriptions of these images (Exp. 1), had difficulties classifying the images as unusual (Exp. 2), and even had lower sensitivity to detect these images in noise than to detect their more probable counterparts (Exp. 3). Taken together, these results place a limit on our abilities for rapid scene perception and suggest that perception is facilitated by prior visual experience.

Global image properties do not guide visual search

While basic visual features such as color, motion, and orientation can guide attention, it is likely that additional features guide search for objects in real-world scenes. Recent work has shown that human observers efficiently extract global scene properties such as mean depth or navigability from a brief glance at a single scene (M. R. Greene & A. Oliva, 2009a, 2009b). Can human observers also efficiently search for an image possessing a particular global scene property among other images lacking that property? Observers searched for scene image targets defined by global properties of naturalness, transience, navigability, and mean depth. All produced inefficient search. Search efficiency for a property was not correlated with its classification threshold time from M. R. Greene and A. Oliva (2009b). Differences in search efficiency between properties can be partially explained by low-level visual features that are correlated with the global property. Overall, while global scene properties can be rapidly classified from a single image, it does not appear to be possible to use those properties to guide attention to one of several images.

Basic level category structure emerges gradually across human ventral visual cortex

Objects can be simultaneously categorized at multiple levels of specificity ranging from very broad (“natural object”) to very distinct (“Mr. Woof”), with a mid-level of generality (basic level: “dog”) often providing the most cognitively useful distinction between categories. It is unknown, however, how this hierarchical representation is achieved in the brain. Using multivoxel pattern analyses, we examined how well each taxonomic level (superordinate, basic, and subordinate) of real-world object categories is represented across occipitotemporal cortex. We found that, although in early visual cortex objects are best represented at the subordinate level (an effect mostly driven by low-level feature overlap between objects in the same category), this advantage diminishes compared to the basic level as we move up the visual hierarchy, disappearing in object-selective regions of occipitotemporal cortex. This pattern stems from a combined increase in within-category similarity (category cohesion) and between-category dissimilarity (category distinctiveness) of neural activity patterns at the basic level, relative to both subordinate and superordinate levels, suggesting that successive visual areas may be optimizing basic level representations.