John K. Tsotsos

Iuliia Kotseruba, John K. Tsotsos · utoronto

To further advance driver monitoring and assistance systems, it is important to understand how drivers allocate their attention, in other words, where do they tend to look and why. Traditionally, factors affecting human visual attention have been divided into bottom-up (involuntary attraction to salient regions) and top-down (driven by the demands of the task being performed). Although both play a role in directing drivers' gaze, most of the existing models for drivers' gaze prediction apply techniques developed for bottom-up saliency and do not consider influences of the drivers' actions explicitly. Likewise, common driving attention benchmarks lack relevant annotations for drivers' actions and the context in which they are performed. Therefore, to enable analysis and modeling of these factors for drivers' gaze prediction, we propose the following: 1) we correct the data processing pipeline used in DR(eye)VE to reduce noise in the recorded gaze data; 2) we then add per-frame labels for driving task and context; 3) we benchmark a number of baseline and SOTA models for saliency and driver gaze prediction and use new annotations to analyze how their performance changes in scenarios involving different tasks; and, lastly, 4) we develop a novel model that modulates drivers' gaze prediction with explicit action and context information. While reducing noise in the DR(eye)VE gaze data improves results of all models, we show that using task information in our proposed model boosts performance even further compared to bottom-up models on the cleaned up data, both overall (by 24% KLD and 89% NSS) and on scenarios that involve performing safety-critical maneuvers and crossing intersections (by up to 10--30% KLD). Extended annotations and code are available at https://github.com/ykotseruba/SCOUT.

Paria Mehrani, John K. Tsotsos · utoronto

Recently, a considerable number of studies in computer vision involves deep neural architectures called vision transformers. Visual processing in these models incorporates computational models that are claimed to implement attention mechanisms. Despite an increasing body of work that attempts to understand the role of attention mechanisms in vision transformers, their effect is largely unknown. Here, we asked if the attention mechanisms in vision transformers exhibit similar effects as those known in human visual attention. To answer this question, we revisited the attention formulation in these models and found that despite the name, computationally, these models perform a special class of relaxation labeling with similarity grouping effects. Additionally, whereas modern experimental findings reveal that human visual attention involves both feed-forward and feedback mechanisms, the purely feed-forward architecture of vision transformers suggests that attention in these models will not have the same effects as those known in humans. To quantify these observations, we evaluated grouping performance in a family of vision transformers. Our results suggest that self-attention modules group figures in the stimuli based on similarity in visual features such as color. Also, in a singleton detection experiment as an instance of saliency detection, we studied if these models exhibit similar effects as those of feed-forward visual salience mechanisms utilized in human visual attention. We found that generally, the transformer-based attention modules assign more salience either to distractors or the ground. Together, our study suggests that the attention mechanisms in vision transformers perform similarity grouping and not attention.

Markus D. Solbach, John K. Tsotsos · utoronto

Our understanding of how visual systems detect, analyze and interpret visual stimuli has advanced greatly. However, the visual systems of all animals do much more; they enable visual behaviours. How well the visual system performs while interacting with the visual environment and how vision is used in the real world have not been well studied, especially in humans. It has been suggested that comparison is the most primitive of psychophysical tasks. Thus, as a probe into these active visual behaviours, we use a same-different task: are two physical 3D objects visually the same? This task seems to be a fundamental cognitive ability. We pose this question to human subjects who are free to move about and examine two real objects in an actual 3D space. Past work has dealt solely with a 2D static version of this problem. We have collected detailed, first-of-its-kind data of humans performing a visuospatial task in hundreds of trials. Strikingly, humans are remarkably good at this task without any training, with a mean accuracy of 93.82%. No learning effect was observed on accuracy after many trials, but some effect was seen for response time, number of fixations and extent of head movement. Subjects demonstrated a variety of complex strategies involving a range of movement and eye fixation changes, suggesting that solutions were developed dynamically and tailored to the specific task.

Josh Goldman, John K. Tsotsos · utoronto

Deep neural networks have achieved impressive performance on many computer vision benchmarks in recent years. However, can we be confident that impressive performance on benchmarks will translate to strong performance in real-world environments? Many environments in the real world are safety critical, and even slight model failures can be catastrophic. Therefore, it is crucial to test models rigorously before deployment. We argue, through both statistical theory and empirical evidence, that selecting representative image datasets for testing a model is likely implausible in many domains. Furthermore, performance statistics calculated with non-representative image datasets are highly unreliable. As a consequence, we cannot guarantee that models which perform well on withheld test images will also perform well in the real world. Creating larger and larger datasets will not help, and bias aware datasets cannot solve this problem either. Ultimately, there is little statistical foundation for evaluating models using withheld test sets. We recommend that future evaluation methodologies focus on assessing a model's decision-making process, rather than metrics such as accuracy.

Iuliia Kotseruba, John K. Tsotsos · utoronto

Accurate prediction of drivers' gaze is an important component of vision-based driver monitoring and assistive systems. Of particular interest are safety-critical episodes, such as performing maneuvers or crossing intersections. In such scenarios, drivers' gaze distribution changes significantly and becomes difficult to predict, especially if the task and context information is represented implicitly, as is common in many state-of-the-art models. However, explicit modeling of top-down factors affecting drivers' attention often requires additional information and annotations that may not be readily available. In this paper, we address the challenge of effective modeling of task and context with common sources of data for use in practical systems. To this end, we introduce SCOUT+, a task- and context-aware model for drivers' gaze prediction, which leverages route and map information inferred from commonly available GPS data. We evaluate our model on two datasets, DR(eye)VE and BDD-A, and demonstrate that using maps improves results compared to bottom-up models and reaches performance comparable to the top-down model SCOUT which relies on privileged ground truth information. Code is available at https://github.com/ykotseruba/SCOUT.

Iuliia Kotseruba, John K. Tsotsos · utoronto

Driving is a visuomotor task, i.e., there is a connection between what drivers see and what they do. While some models of drivers' gaze account for top-down effects of drivers' actions, the majority learn only bottom-up correlations between human gaze and driving footage. The crux of the problem is lack of public data with annotations that could be used to train top-down models and evaluate how well models of any kind capture effects of task on attention. As a result, top-down models are trained and evaluated on private data and public benchmarks measure only the overall fit to human data. In this paper, we focus on data limitations by examining four large-scale public datasets, DR(eye)VE, BDD-A, MAAD, and LBW, used to train and evaluate algorithms for drivers' gaze prediction. We define a set of driving tasks (lateral and longitudinal maneuvers) and context elements (intersections and right-of-way) known to affect drivers' attention, augment the datasets with annotations based on the said definitions, and analyze the characteristics of data recording and processing pipelines w.r.t. capturing what the drivers see and do. In sum, the contributions of this work are: 1) quantifying biases of the public datasets, 2) examining performance of the SOTA bottom-up models on subsets of the data involving non-trivial drivers' actions, 3) linking shortcomings of the bottom-up models to data limitations, and 4) recommendations for future data collection and processing. The new annotations and code for reproducing the results is available at https://github.com/ykotseruba/SCOUT.

Iuliia Kotseruba, John K. Tsotsos · utoronto

Generalization of deep-learning-based (DL) computer vision algorithms to various image perturbations is hard to establish and remains an active area of research. The majority of past analyses focused on the images already captured, whereas effects of the image formation pipeline and environment are less studied. In this paper, we address this issue by analyzing the impact of capture conditions, such as camera parameters and lighting, on DL model performance on 3 vision tasks -- image classification, object detection, and visual question answering (VQA). To this end, we assess capture bias in common vision datasets and create a new benchmark, SNAP (for $\textbf{S}$hutter speed, ISO se$\textbf{N}$sitivity, and $\textbf{AP}$erture), consisting of images of objects taken under controlled lighting conditions and with densely sampled camera settings. We then evaluate a large number of DL vision models and show the effects of capture conditions on each selected vision task. Lastly, we conduct an experiment to establish a human baseline for the VQA task. Our results show that computer vision datasets are significantly biased, the models trained on this data do not reach human accuracy even on the well-exposed images, and are susceptible to both major exposure changes and minute variations of camera settings. Code and data can be found at https://github.com/ykotseruba/SNAP

Iuliia Kotseruba, John K. Tsotsos

Driving is a routine activity for many, but it is far from simple. Drivers deal with multiple concurrent tasks, such as keeping the vehicle in the lane, observing and anticipating the actions of other road users, reacting to hazards, and dealing with distractions inside and outside the vehicle. Failure to notice and respond to the surrounding objects and events can cause accidents. The ongoing improvements of the road infrastructure and vehicle mechanical design have made driving safer overall. Nevertheless, the problem of driver inattention has remained one of the primary causes of accidents. Therefore, understanding where the drivers look and why they do so can help eliminate sources of distractions and identify unsafe attention patterns. Research on driver attention has implications for many practical applications such as policy-making, improving driver education, enhancing road infrastructure and in-vehicle infotainment systems, as well as designing systems for driver monitoring, driver assistance, and automated driving. This report covers the literature on changes in drivers' visual attention distribution due to factors, internal and external to the driver. Aspects of attention during driving have been explored across multiple disciplines, including psychology, human factors, human-computer interaction, intelligent transportation, and computer vision, each offering different perspectives, goals, and explanations for the observed phenomena. We link cross-disciplinary theoretical and behavioral research on driver's attention to practical solutions. Furthermore, limitations and directions for future research are discussed. This report is based on over 175 behavioral studies, nearly 100 practical papers, 20 datasets, and over 70 surveys published since 2010. A curated list of papers used for this report is available at \url{https://github.com/ykotseruba/attention_and_driving}.

Bao Xin Chen, John K. Tsotsos

In this paper, we demonstrate a novel algorithm that uses ellipse fitting to estimate the bounding box rotation angle and size with the segmentation(mask) on the target for online and real-time visual object tracking. Our method, SiamMask_E, improves the bounding box fitting procedure of the state-of-the-art object tracking algorithm SiamMask and still retains a fast-tracking frame rate (80 fps) on a system equipped with GPU (GeForce GTX 1080 Ti or higher). We tested our approach on the visual object tracking datasets (VOT2016, VOT2018, and VOT2019) that were labeled with rotated bounding boxes. By comparing with the original SiamMask, we achieved an improved Accuracy of 0.652 and 0.309 EAO on VOT2019, which is 0.056 and 0.026 higher than the original SiamMask. The implementation is available on GitHub: https://github.com/baoxinchen/siammask_e.

Calden Wloka, Toni Kunić, Iuliia Kotseruba et al.

The Saliency Model Implementation Library for Experimental Research (SMILER) is a new software package which provides an open, standardized, and extensible framework for maintaining and executing computational saliency models. This work drastically reduces the human effort required to apply saliency algorithms to new tasks and datasets, while also ensuring consistency and procedural correctness for results and conclusions produced by different parties. At its launch SMILER already includes twenty three saliency models (fourteen models based in MATLAB and nine supported through containerization), and the open design of SMILER encourages this number to grow with future contributions from the community. The project may be downloaded and contributed to through its GitHub page: https://github.com/tsotsoslab/smiler

Paria Mehrani, John K. Tsotsos

The mechanisms involved in transforming early visual signals to curvature representations in V4 are unknown. We propose a hierarchical model that reveals V1/V2 encodings that are essential components for this transformation to the reported curvature representations in V4. Then, by relaxing the often-imposed prior of a single Gaussian, V4 shape selectivity is learned in the last layer of the hierarchy from Macaque V4 responses. We found that V4 cells integrate multiple shape parts from the full spatial extent of their receptive fields with similar excitatory and inhibitory contributions. Our results uncover new details in existing data about shape selectivity in V4 neurons that with further experiments can enhance our understanding of processing in this area. Accordingly, we propose designs for a stimulus set that allow removing shape parts without disturbing the curvature signal to isolate part contributions to V4 responses.

Christian Korbach, Markus D. Solbach, Raphael Memmesheimer et al.

The presentation and analysis of image data from a single viewpoint are often not sufficient to solve a task. Several viewpoints are necessary to obtain more information. The next-best-view problem attempts to find the optimal viewpoint with the greatest information gain for the underlying task. In this work, a robot arm holds an object in its end-effector and searches for a sequence of next-best-view to explicitly identify the object. We use Soft Actor-Critic (SAC), a method of deep reinforcement learning, to learn these next-best-views for a specific set of objects. The evaluation shows that an agent can learn to determine an object pose to which the robot arm should move an object. This leads to a viewpoint that provides a more accurate prediction to distinguish such an object from other objects better. We make the code publicly available for the scientific community and for reproducibility.

Markus D. Solbach, John K. Tsotsos

The STAR architecture was designed to test the value of the full Selective Tuning model of visual attention for complex real-world visuospatial tasks and behaviors. However, knowledge of how humans solve such tasks in 3D as active observers is lean. We thus devised a novel experimental setup and examined such behavior. We discovered that humans exhibit a variety of problem-solving strategies whose breadth and complexity are surprising and not easily handled by current methodologies. It is apparent that solution methods are dynamically composed by hypothesizing sequences of actions, testing them, and if they fail, trying different ones. The importance of active observation is striking as is the lack of any learning effect. These results inform our Cognitive Program representation of STAR extending its relevance to real-world tasks.

Iuliia Kotseruba, Manos Papagelis, John K. Tsotsos

This work aims to study the dynamic between research in the industry and academia in computer vision. The results are demonstrated on a set of top-5 vision conferences that are representative of the field. Since data for such analysis was not readily available, significant effort was spent on gathering and processing meta-data from the original publications. First, this study quantifies the share of industry-sponsored research. Specifically, it shows that the proportion of papers published by industry-affiliated researchers is increasing and that more academics join companies or collaborate with them. Next, the possible impact of industry presence is further explored, namely in the distribution of research topics and citation patterns. The results indicate that the distribution of the research topics is similar in industry and academic papers. However, there is a strong preference towards citing industry papers. Finally, possible reasons for citation bias, such as code availability and influence, are investigated.

John K. Tsotsos, Jun Luo

Learned systems in the domain of visual recognition and cognition impress in part because even though they are trained with datasets many orders of magnitude smaller than the full population of possible images, they exhibit sufficient generalization to be applicable to new and previously unseen data. Since training data sets typically represent small sampling of a domain, the possibility of bias in their composition is very real. But what are the limits of generalization given such bias, and up to what point might it be sufficient for a real problem task? Although many have examined issues regarding generalization, this question may require examining the data itself. Here, we focus on the characteristics of the training data that may play a role. Other disciplines have grappled with these problems, most interestingly epidemiology, where experimental bias is a critical concern. The range and nature of data biases seen clinically are really quite relatable to learned vision systems. One obvious way to deal with bias is to ensure a large enough training set, but this might be infeasible for many domains. Another approach might be to perform a statistical analysis of the actual training set, to determine if all aspects of the domain are fairly captured. This too is difficult, in part because the full set of variables might not be known, or perhaps not even knowable. Here, we try a different approach in the tradition of the Thought Experiment, whose most famous instance may be Schrödinger's Cat. There are many types of bias as will be seen, but we focus only on one, selection bias. The point of the thought experiment is not to demonstrate problems with all learned systems. Rather, this might be a simple theoretical tool to probe into bias during data collection to highlight deficiencies that might then deserve extra attention either in data collection or system development.

Markus D. Solbach, John K. Tsotsos

Despite the recent successes in computer vision, there remain new avenues to explore. In this work, we propose a new dataset to investigate the effect of self-occlusion on deep neural networks. With TEOS (The Effect of Self-Occlusion), we propose a 3D blocks world dataset that focuses on the geometric shape of 3D objects and their omnipresent challenge of self-occlusion. We designed TEOS to investigate the role of self-occlusion in the context of object classification. Even though remarkable progress has been seen in object classification, self-occlusion is a challenge. In the real-world, self-occlusion of 3D objects still presents significant challenges for deep learning approaches. However, humans deal with this by deploying complex strategies, for instance, by changing the viewpoint or manipulating the scene to gather necessary information. With TEOS, we present a dataset of two difficulty levels (L1 and L2 ), containing 36 and 12 objects, respectively. We provide 738 uniformly sampled views of each object, their mask, object and camera position, orientation, amount of self-occlusion, as well as the CAD model of each object. We present baseline evaluations with five well-known classification deep neural networks and show that TEOS poses a significant challenge for all of them. The dataset, as well as the pre-trained models, are made publicly available for the scientific community under https://nvision2.data.eecs.yorku.ca/TEOS.

Mahdi Biparva, David Fernández-Llorca, Rubén Izquierdo-Gonzalo et al.

In highway scenarios, an alert human driver will typically anticipate early cut-in/cut-out maneuvers of surrounding vehicles using visual cues mainly. Autonomous vehicles must anticipate these situations at an early stage too, to increase their safety and efficiency. In this work, lane-change recognition and prediction tasks are posed as video action recognition problems. Up to four different two-stream-based approaches, that have been successfully applied to address human action recognition, are adapted here by stacking visual cues from forward-looking video cameras to recognize and anticipate lane-changes of target vehicles. We study the influence of context and observation horizons on performance, and different prediction horizons are analyzed. The different models are trained and evaluated using the PREVENTION dataset. The obtained results clearly demonstrate the potential of these methodologies to serve as robust predictors of future lane-changes of surrounding vehicles proving an accuracy higher than 90% in time horizons of between 1-2 seconds.

John K. Tsotsos, Omar Abid, Iuliia Kotseruba et al.

The study of attentional processing in vision has a long and deep history. Recently, several papers have presented insightful perspectives into how the coordination of multiple attentional functions in the brain might occur. These begin with experimental observations and the authors propose structures, processes, and computations that might explain those observations. Here, we consider a perspective that past works have not, as a complementary approach to the experimentally-grounded ones. We approach the same problem as past authors but from the other end of the computational spectrum, from the problem nature, as Marr's Computational Level would prescribe. What problem must the brain solve when orchestrating attentional processes in order to successfully complete one of the myriad possible visuospatial tasks at which we as humans excel? The hope, of course, is for the approaches to eventually meet and thus form a complete theory, but this is likely not soon. We make the first steps towards this by addressing the necessity of attentional control, examining the breadth and computational difficulty of the visuospatial and attentional tasks seen in human behavior, and suggesting a sketch of how attentional control might arise in the brain. The key conclusions of this paper are that an executive controller is necessary for human attentional function in vision, and that there is a 'first principles' computational approach to its understanding that is complementary to the previous approaches that focus on modelling or learning from experimental observations directly.

Calden Wloka, John K. Tsotsos

When applying a convolutional kernel to an image, if the output is to remain the same size as the input then some form of padding is required around the image boundary, meaning that for each layer of convolution in a convolutional neural network (CNN), a strip of pixels equal to the half-width of the kernel size is produced with a non-veridical representation. Although most CNN kernels are small to reduce the parameter load of a network, this non-veridical area compounds with each convolutional layer. The tendency toward deeper and deeper networks combined with stride-based down-sampling means that the propagation of this region can end up covering a non-negligable portion of the image. Although this issue with convolutions has been well acknowledged over the years, the impact of this degraded peripheral representation on modern network behavior has not been fully quantified. What are the limits of translation invariance? Does image padding successfully mitigate the issue, or is performance affected as an object moves between the image border and center? Using Mask R-CNN as an experimental model, we design a dataset and methodology to quantify the spatial dependency of network performance. Our dataset is constructed by inserting objects into high resolution backgrounds, thereby allowing us to crop sub-images which place target objects at specific locations relative to the image border. By probing the behaviour of Mask R-CNN across a selection of target locations, we see clear patterns of performance degredation near the image boundary, and in particular in the image corners. Quantifying both the extent and magnitude of this spatial anisotropy in network performance is important for the deployment of deep networks into unconstrained and realistic environments in which the location of objects or regions of interest are not guaranteed to be well localized within a given image.

Mahdi Biparva, John Tsotsos

Feature disentanglement of the foreground target objects and the background surrounding context has not been yet fully accomplished. The lack of network interpretability prevents advancing for feature disentanglement and better generalization robustness. We study the role of the context on interfering with a disentangled foreground target object representation in this work. We hypothesize that the representation of the surrounding context is heavily tied with the foreground object due to the dense hierarchical parametrization of convolutional networks with under-constrained learning algorithms. Working on a framework that benefits from the bottom-up and top-down processing paradigms, we investigate a systematic approach to shift learned representations in feedforward networks from the emphasis on the irrelevant context to the foreground objects. The top-down processing provides importance maps as the means of the network internal self-interpretation that will guide the learning algorithm to focus on the relevant foreground regions towards achieving a more robust representations. We define an experimental evaluation setup with the role of context emphasized using the MNIST dataset. The experimental results reveal not only that the label prediction accuracy is improved but also a higher degree of robustness to the background perturbation using various noise generation methods is obtained.

Markus D. Solbach, John K. Tsotsos

Most past and present research in computer vision involves passively observed data. Humans, however, are active observers outside the lab; they explore, search, select what and how to look. Nonetheless, how exactly active observation occurs in humans so that it can inform the design of active computer vision systems is an open problem. PESAO is designed for investigating active, visual observation in a 3D world. The goal was to build an experimental setup for various active perception tasks with human subjects (active observers) in mind that is capable of tracking the head and gaze. While many studies explore human performances, usually, they use line drawings portrayed in 2D, and no active observer is involved. PESAO allows us to bring many studies to the three-dimensional world, even involving active observers. In our instantiation, it spans an area of 400cm x 300cm and can track active observers at a frequency of 120Hz. Furthermore, PESAO provides tracking and recording of 6D head motion, gaze, eye movement-type, first-person video, head-mounted IMU sensor, birds-eye video, and experimenter notes. All are synchronized at microsecond resolution.

David Fernández-Llorca, Mahdi Biparva, Rubén Izquierdo-Gonzalo et al.

In highway scenarios, an alert human driver will typically anticipate early cut-in and cut-out maneuvers of surrounding vehicles using only visual cues. An automated system must anticipate these situations at an early stage too, to increase the safety and the efficiency of its performance. To deal with lane-change recognition and prediction of surrounding vehicles, we pose the problem as an action recognition/prediction problem by stacking visual cues from video cameras. Two video action recognition approaches are analyzed: two-stream convolutional networks and spatiotemporal multiplier networks. Different sizes of the regions around the vehicles are analyzed, evaluating the importance of the interaction between vehicles and the context information in the performance. In addition, different prediction horizons are evaluated. The obtained results demonstrate the potential of these methodologies to serve as robust predictors of future lane-changes of surrounding vehicles in time horizons between 1 and 2 seconds.

Iuliia Kotseruba, Calden Wloka, Amir Rasouli et al.

Recent advances in the field of saliency have concentrated on fixation prediction, with benchmarks reaching saturation. However, there is an extensive body of works in psychology and neuroscience that describe aspects of human visual attention that might not be adequately captured by current approaches. Here, we investigate singleton detection, which can be thought of as a canonical example of salience. We introduce two novel datasets, one with psychophysical patterns and one with natural odd-one-out stimuli. Using these datasets we demonstrate through extensive experimentation that nearly all saliency algorithms do not adequately respond to singleton targets in synthetic and natural images. Furthermore, we investigate the effect of training state-of-the-art CNN-based saliency models on these types of stimuli and conclude that the additional training data does not lead to a significant improvement of their ability to find odd-one-out targets. Datasets are available at http://data.nvision2.eecs.yorku.ca/P3O3/.

Amir Rasouli, Iuliia Kotseruba, John K. Tsotsos

One of the major challenges for autonomous vehicles in urban environments is to understand and predict other road users' actions, in particular, pedestrians at the point of crossing. The common approach to solving this problem is to use the motion history of the agents to predict their future trajectories. However, pedestrians exhibit highly variable actions most of which cannot be understood without visual observation of the pedestrians themselves and their surroundings. To this end, we propose a solution for the problem of pedestrian action anticipation at the point of crossing. Our approach uses a novel stacked RNN architecture in which information collected from various sources, both scene dynamics and visual features, is gradually fused into the network at different levels of processing. We show, via extensive empirical evaluations, that the proposed algorithm achieves a higher prediction accuracy compared to alternative recurrent network architectures. We conduct experiments to investigate the impact of the length of observation, time to event and types of features on the performance of the proposed method. Finally, we demonstrate how different data fusion strategies impact prediction accuracy.

Mahdi Biparva, John Tsotsos

Deep neural networks have evolved to become power demanding and consequently difficult to apply to small-size mobile platforms. Network parameter reduction methods have been introduced to systematically deal with the computational and memory complexity of deep networks. We propose to examine the ability of attentive connection pruning to deal with redundancy reduction in neural networks as a contribution to the reduction of computational demand. In this work, we describe a Top-Down attention mechanism that is added to a Bottom-Up feedforward network to select important connections and subsequently prune redundant ones at all parametric layers. Our method not only introduces a novel hierarchical selection mechanism as the basis of pruning but also remains competitive with previous baseline methods in the experimental evaluation. We conduct experiments using different network architectures on popular benchmark datasets to show high compression ratio is achievable with negligible loss of accuracy.

Mahdi Biparva, John Tsotsos

Convolutional neural networks model the transformation of the input sensory data at the bottom of a network hierarchy to the semantic information at the top of the visual hierarchy. Feedforward processing is sufficient for some object recognition tasks. Top-Down selection is potentially required in addition to the Bottom-Up feedforward pass. It can, in part, address the shortcoming of the loss of location information imposed by the hierarchical feature pyramids. We propose a unified 2-pass framework for object segmentation that augments Bottom-Up \convnets with a Top-Down selection network. We utilize the top-down selection gating activities to modulate the bottom-up hidden activities for segmentation predictions. We develop an end-to-end multi-task framework with loss terms satisfying task requirements at the two ends of the network. We evaluate the proposed network on benchmark datasets for semantic segmentation, and show that networks with the Top-Down selection capability outperform the baseline model. Additionally, we shed light on the superior aspects of the new segmentation paradigm and qualitatively and quantitatively support the efficiency of the novel framework over the baseline model that relies purely on parametric skip connections.

John K. Tsotsos, Iuliia Kotseruba, Alexander Andreopoulos et al.

Why do some continue to wonder about the success and dominance of deep learning methods in computer vision and AI? Is it not enough that these methods provide practical solutions to many problems? Well no, it is not enough, at least for those who feel there should be a science that underpins all of this and that we should have a clear understanding of how this success was achieved. Here, this paper proposes that the dominance we are witnessing would not have been possible by the methods of deep learning alone: the tacit change has been the evolution of empirical practice in computer vision and AI over the past decades. We demonstrate this by examining the distribution of sensor settings in vision datasets and performance of both classic and deep learning algorithms under various camera settings. This reveals a strong mismatch between optimal performance ranges of classical theory-driven algorithms and sensor setting distributions in the common vision datasets, while data-driven models were trained for those datasets. The head-to-head comparisons between data-driven and theory-driven models were therefore unknowingly biased against the theory-driven models.

Bao Xin Chen, Raghavender Sahdev, Dekun Wu et al.

Scene Classification has been addressed with numerous techniques in computer vision literature. However, with the increasing number of scene classes in datasets in the field, it has become difficult to achieve high accuracy in the context of robotics. In this paper, we implement an approach which combines traditional deep learning techniques with natural language processing methods to generate a word embedding based Scene Classification algorithm. We use the key idea that context (objects in the scene) of an image should be representative of the scene label meaning a group of objects could assist to predict the scene class. Objects present in the scene are represented by vectors and the images are re-classified based on the objects present in the scene to refine the initial classification by a Convolutional Neural Network (CNN). In our approach we address indoor Scene Classification task using a model trained with a reduced pre-processed version of the Places365 dataset and an empirical analysis is done on a real-world dataset that we built by capturing image sequences using a GoPro camera. We also report results obtained on a subset of the Places365 dataset using our approach and additionally show a deployment of our approach on a robot operating in a real-world environment.

Paria Mehrani, Andrei Mouraviev, John K. Tsotsos

There is still much to understand about the color processing mechanisms in the brain and the transformation from cone-opponent representations to perceptual hues. Moreover, it is unclear which areas(s) in the brain represent unique hues. We propose a hierarchical model inspired by the neuronal mechanisms in the brain for local hue representation, which reveals the contributions of each visual cortical area in hue representation. Local hue encoding is achieved through incrementally increasing processing nonlinearities beginning with cone input. Besides employing nonlinear rectifications, we propose multiplicative modulations as a form of nonlinearity. Our simulation results indicate that multiplicative modulations have significant contributions in encoding of hues along intermediate directions in the MacLeod-Boynton diagram and that model V4 neurons have the capacity to encode unique hues. Additionally, responses of our model neurons resemble those of biological color cells, suggesting that our model provides a novel formulation of the brain's color processing pathway.

Amir Rosenfeld, Richard Zemel, John K. Tsotsos

Predicting human perceptual similarity is a challenging subject of ongoing research. The visual process underlying this aspect of human vision is thought to employ multiple different levels of visual analysis (shapes, objects, texture, layout, color, etc). In this paper, we postulate that the perception of image similarity is not an explicitly learned capability, but rather one that is a byproduct of learning others. This claim is supported by leveraging representations learned from a diverse set of visual tasks and using them jointly to predict perceptual similarity. This is done via simple feature concatenation, without any further learning. Nevertheless, experiments performed on the challenging Totally-Looks-Like (TLL) benchmark significantly surpass recent baselines, closing much of the reported gap towards prediction of human perceptual similarity. We provide an analysis of these results and discuss them in a broader context of emergent visual capabilities and their implications on the course of machine-vision research.

John K. Tsotsos, Iuliia Kotseruba, Calden Wloka

The current dominant visual processing paradigm in both human and machine research is the feedforward, layered hierarchy of neural-like processing elements. Within this paradigm, visual saliency is seen by many to have a specific role, namely that of early selection. Early selection is thought to enable very fast visual performance by limiting processing to only the most salient candidate portions of an image. This strategy has led to a plethora of saliency algorithms that have indeed improved processing time efficiency in machine algorithms, which in turn have strengthened the suggestion that human vision also employs a similar early selection strategy. However, at least one set of critical tests of this idea has never been performed with respect to the role of early selection in human vision. How would the best of the current saliency models perform on the stimuli used by experimentalists who first provided evidence for this visual processing paradigm? Would the algorithms really provide correct candidate sub-images to enable fast categorization on those same images? Do humans really need this early selection for their impressive performance? Here, we report on a new series of tests of these questions whose results suggest that it is quite unlikely that such an early selection process has any role in human rapid visual categorization.

Paria Mehrani, John K. Tsotsos

The face-vase illusion introduced by Rubin demonstrates how one can switch back and forth between two different interpretations depending on how the figure outlines are assigned [1]. This border ownership assignment is an important step in the perception of forms. Zhou et al. [2] found neurons in the visual cortex whose responses not only depend on the local features present in their classical receptive fields, but also on their contextual information. Various models proposed that feedback from higher ventral areas or lateral connections could provide the required contextual information. However, some studies [3, 4, 5] ruled out the plausibility of models exclusively based on lateral connections. In addition, further evidence [6] suggests that ventral feedback even from V4 is not fast enough to provide context to border ownership neurons in either V1 or V2. As a result, the border ownership assignment mechanism in the brain is a mystery yet to be solved. Here, we test with computational simulations the hypothesis that the dorsal stream provides the global information to border ownership cells in the ventral stream. Our proposed model incorporates early recurrence from the dorsal pathway as well as lateral modulations within the ventral stream. Our simulation experiments show that our model border ownership neurons, similar to their biological counterparts, exhibit different responses to figures on either side of the border.

Audrey M. B. Wong-Kee-You, John K. Tsotsos, Scott A. Adler

The capacity to filter out irrelevant information from our environment is critical to efficient processing. Yet, during development, when building a knowledge base of the world is occurring, the ability to selectively allocate attentional resources is limited (e.g., Amso & Scerif, 2015). In adulthood, research has demonstrated that surrounding the spatial location of attentional focus is a suppressive field, resulting from top-down attention promoting the processing of relevant stimuli and inhibiting surrounding distractors (e.g., Hopf et al., 2006). It is not fully known, however, whether this phenomenon manifests in development. In the current study, we examined whether spatial suppression surrounding the focus of visual attention is exhibited in developmental age groups. Participants between 12 and 27 years of age exhibited spatial suppression surrounding their focus of visual attention. Their accuracy increased as a function of the separation distance between a spatially cued (and attended) target and a second target, suggesting that a ring of suppression surrounded the attended target. When a central cue was instead presented and therefore attention was no longer spatially cued, surround suppression was not observed, indicating that our initial findings of suppression were indeed related to the focus of attention. Attentional surround suppression was not observed in 8- to 11-years-olds, even with a longer spatial cue presentation time, demonstrating that the lack of the effect at these ages is not due to slowed attentional feedback processes. Our findings demonstrate that top-down attentional processes are still immature until approximately 12 years of age, and that they continue to be refined throughout adolescence, converging well with previous research on attentional development.

Amir Rosenfeld, Richard Zemel, John K. Tsotsos

We showcase a family of common failures of state-of-the art object detectors. These are obtained by replacing image sub-regions by another sub-image that contains a trained object. We call this "object transplanting". Modifying an image in this manner is shown to have a non-local impact on object detection. Slight changes in object position can affect its identity according to an object detector as well as that of other objects in the image. We provide some analysis and suggest possible reasons for the reported phenomena.

John K. Tsotsos, Iuliia Kotseruba, Amir Rasouli et al.

It is almost universal to regard attention as the facility that permits an agent, human or machine, to give priority processing resources to relevant stimuli while ignoring the irrelevant. The reality of how this might manifest itself throughout all the forms of perceptual and cognitive processes possessed by humans, however, is not as clear. Here we examine this reality with a broad perspective in order to highlight the myriad ways that attentional processes impact both perception and cognition. The paper concludes by showing two real world problems that exhibit sufficient complexity to illustrate the ways in which attention and cognition connect. These then point to new avenues of research that might illuminate the overall cognitive architecture of spatial cognition.

Amir Rasouli, John K. Tsotsos

One of the major challenges that autonomous cars are facing today is driving in urban environments. To make it a reality, autonomous vehicles require the ability to communicate with other road users and understand their intentions. Such interactions are essential between the vehicles and pedestrians as the most vulnerable road users. Understanding pedestrian behavior, however, is not intuitive and depends on various factors such as demographics of the pedestrians, traffic dynamics, environmental conditions, etc. In this paper, we identify these factors by surveying pedestrian behavior studies, both the classical works on pedestrian-driver interaction and the modern ones that involve autonomous vehicles. To this end, we will discuss various methods of studying pedestrian behavior, and analyze how the factors identified in the literature are interrelated. We will also review the practical applications aimed at solving the interaction problem including design approaches for autonomous vehicles that communicate with pedestrians and visual perception and reasoning algorithms tailored to understanding pedestrian intention. Based on our findings, we will discuss the open problems and propose future research directions.

Markus D. Solbach, Stephen Voland, Jeff Edmonds et al.

We present a Polyhedral Scene Generator system which creates a random scene based on a few user parameters, renders the scene from random view points and creates a dataset containing the renderings and corresponding annotation files. We hope that this generator will enable research on how a program could parse a scene if it had multiple viewpoints to consider. For ambiguous scenes, typically people move their head or change their position to see the scene from different angles as well as seeing how it changes while they move; this research field is called active perception. The random scene generator presented is designed to support research in this field by generating images of scenes with known complexity characteristics and with verifiable properties with respect to the distribution of features across a population. Thus, it is well-suited for research in active perception without the requirement of a live 3D environment and mobile sensing agent, including comparative performance evaluations. The system is publicly available at https://polyhedral.eecs.yorku.ca.

Amir Rosenfeld, Markus D. Solbach, John K. Tsotsos

Perceptual judgment of image similarity by humans relies on rich internal representations ranging from low-level features to high-level concepts, scene properties and even cultural associations. However, existing methods and datasets attempting to explain perceived similarity use stimuli which arguably do not cover the full breadth of factors that affect human similarity judgments, even those geared toward this goal. We introduce a new dataset dubbed Totally-Looks-Like (TLL) after a popular entertainment website, which contains images paired by humans as being visually similar. The dataset contains 6016 image-pairs from the wild, shedding light upon a rich and diverse set of criteria employed by human beings. We conduct experiments to try to reproduce the pairings via features extracted from state-of-the-art deep convolutional neural networks, as well as additional human experiments to verify the consistency of the collected data. Though we create conditions to artificially make the matching task increasingly easier, we show that machine-extracted representations perform very poorly in terms of reproducing the matching selected by humans. We discuss and analyze these results, suggesting future directions for improvement of learned image representations.

Amir Rosenfeld, John K. Tsotsos

While great advances are made in pattern recognition and machine learning, the successes of such fields remain restricted to narrow applications and seem to break down when training data is scarce, a shift in domain occurs, or when intelligent reasoning is required for rapid adaptation to new environments. In this work, we list several of the shortcomings of modern machine-learning solutions, specifically in the contexts of computer vision and in reinforcement learning and suggest directions to explore in order to try to ameliorate these weaknesses.

Amir Rosenfeld, John K. Tsotsos

There is no denying the tremendous leap in the performance of machine learning methods in the past half-decade. Some might even say that specific sub-fields in pattern recognition, such as machine-vision, are as good as solved, reaching human and super-human levels. Arguably, lack of training data and computation power are all that stand between us and solving the remaining ones. In this position paper we underline cases in vision which are challenging to machines and even to human observers. This is to show limitations of contemporary models that are hard to ameliorate by following the current trend to increase training data, network capacity or computational power. Moreover, we claim that attempting to do so is in principle a suboptimal approach. We provide a taster of such examples in hope to encourage and challenge the machine learning community to develop new directions to solve the said difficulties.

Amir Rasouli, John K. Tsotsos

Today, one of the major challenges that autonomous vehicles are facing is the ability to drive in urban environments. Such a task requires communication between autonomous vehicles and other road users in order to resolve various traffic ambiguities. The interaction between road users is a form of negotiation in which the parties involved have to share their attention regarding a common objective or a goal (e.g. crossing an intersection), and coordinate their actions in order to accomplish it. In this literature review we aim to address the interaction problem between pedestrians and drivers (or vehicles) from joint attention point of view. More specifically, we will discuss the theoretical background behind joint attention, its application to traffic interaction and practical approaches to implementing joint attention for autonomous vehicles.

Amir Rosenfeld, John K. Tsotsos

Training deep neural networks results in strong learned representations that show good generalization capabilities. In most cases, training involves iterative modification of all weights inside the network via back-propagation. In Extreme Learning Machines, it has been suggested to set the first layer of a network to fixed random values instead of learning it. In this paper, we propose to take this approach a step further and fix almost all layers of a deep convolutional neural network, allowing only a small portion of the weights to be learned. As our experiments show, fixing even the majority of the parameters of the network often results in performance which is on par with the performance of learning all of them. The implications of this intriguing property of deep neural networks are discussed and we suggest ways to harness it to create more robust representations.

Calden Wloka, Iuliia Kotseruba, John K. Tsotsos

Visual attention is a field with a considerable history, with eye movement control and prediction forming an important subfield. Fixation modeling in the past decades has been largely dominated computationally by a number of highly influential bottom-up saliency models, such as the Itti-Koch-Niebur model. The accuracy of such models has dramatically increased recently due to deep learning. However, on static images the emphasis of these models has largely been based on non-ordered prediction of fixations through a saliency map. Very few implemented models can generate temporally ordered human-like sequences of saccades beyond an initial fixation point. Towards addressing these shortcomings we present STAR-FC, a novel multi-saccade generator based on a central/peripheral integration of deep learning-based saliency and lower-level feature-based saliency. We have evaluated our model using the CAT2000 database, successfully predicting human patterns of fixation with equivalent accuracy and quality compared to what can be achieved by using one human sequence to predict another. This is a significant improvement over fixation sequences predicted by state-of-the-art saliency algorithms.

Iuliia Kotseruba, John K. Tsotsos

In this paper we present STAR-RT - the first working prototype of Selective Tuning Attention Reference (STAR) model and Cognitive Programs (CPs). The Selective Tuning (ST) model received substantial support through psychological and neurophysiological experiments. The STAR framework expands ST and applies it to practical visual tasks. In order to do so, similarly to many cognitive architectures, STAR combines the visual hierarchy (based on ST) with the executive controller, working and short-term memory components and fixation controller. CPs in turn enable the communication among all these elements for visual task execution. To test the relevance of the system in a realistic context, we implemented the necessary components of STAR and designed CPs for playing two closed-source video games - Canabaltand Robot Unicorn Attack. Since both games run in a browser window, our algorithm has the same amount of information and the same amount of time to react to the events on the screen as a human player would. STAR-RT plays both games in real time using only visual input and achieves scores comparable to human expert players. It thus provides an existence proof for the utility of the particular CP structure and primitives used and the potential for continued experimentation and verification of their utility in broader scenarios.

Amir Rosenfeld, Mahdi Biparva, John K. Tsotsos

Visual priming is known to affect the human visual system to allow detection of scene elements, even those that may have been near unnoticeable before, such as the presence of camouflaged animals. This process has been shown to be an effect of top-down signaling in the visual system triggered by the said cue. In this paper, we propose a mechanism to mimic the process of priming in the context of object detection and segmentation. We view priming as having a modulatory, cue dependent effect on layers of features within a network. Our results show how such a process can be complementary to, and at times more effective than simple post-processing applied to the output of the network, notably so in cases where the object is hard to detect such as in severe noise. Moreover, we find the effects of priming are sometimes stronger when early visual layers are affected. Overall, our experiments confirm that top-down signals can go a long way in improving object detection and segmentation.

Markus D. Solbach, John K. Tsotsos

Falls are serious and costly for elderly people. The Centers for Disease Control and Prevention of the US reports that millions of older people, 65 and older, fall each year at least once. Serious injuries such as; hip fractures, broken bones or head injury, are caused by 20% of the falls. The time it takes to respond and treat a fallen person is crucial. With this paper we present a new , non-invasive system for fallen people detection. Our approach uses only stereo camera data for passively sensing the environment. The key novelty is a human fall detector which uses a CNN based human pose estimator in combination with stereo data to reconstruct the human pose in 3D and estimate the ground plane in 3D. Furthermore, our system consists of a reasoning module which formulates a number of measures to reason whether a person is fallen. We have tested our approach in different scenarios covering most activities elderly people might encounter living at home. Based on our extensive evaluations, our systems shows high accuracy and almost no miss-classification. To reproduce our results, the implementation is publicly available to the scientific community.

Paria Mehrani, Andrei Mouraviev, Oscar J. Avella Gonzalez et al.

A biologically plausible computational model for color representation is introduced. We present a mechanistic hierarchical model of neurons that not only successfully encodes local hue, but also explicitly reveals how the contributions of each visual cortical layer participating in the process can lead to a hue representation. Our proposed model benefits from studies on the visual cortex and builds a network of single-opponent and hue-selective neurons. Local hue encoding is achieved through gradually increasing nonlinearity in terms of cone inputs to single-opponent cells. We demonstrate that our model's single-opponent neurons have wide tuning curves, while the hue-selective neurons in our model V4 layer exhibit narrower tunings, resembling those in V4 of the primate visual system. Our simulation experiments suggest that neurons in V4 or later layers have the capacity of encoding unique hues. Moreover, with a few examples, we present the possibility of spanning the infinite space of physical hues by combining the hue-selective neurons in our model.

Yulong Wu, John Tsotsos

Camera parameters not only play an important role in determining the visual quality of perceived images, but also affect the performance of vision algorithms, for a vision-guided robot. By quantitatively evaluating four object detection algorithms, with respect to varying ambient illumination, shutter speed and voltage gain, it is observed that the performance of the algorithms is highly dependent on these variables. From this observation, a novel active control of camera parameters method is proposed, to make robot vision more robust under different light conditions. Experimental results demonstrate the effectiveness of our proposed approach, which improves the performance of object detection algorithms, compared with the conventional auto-exposure algorithm.

Amir Rosenfeld, John K. Tsotsos

Given an existing trained neural network, it is often desirable to learn new capabilities without hindering performance of those already learned. Existing approaches either learn sub-optimal solutions, require joint training, or incur a substantial increment in the number of parameters for each added domain, typically as many as the original network. We propose a method called \emph{Deep Adaptation Networks} (DAN) that constrains newly learned filters to be linear combinations of existing ones. DANs precisely preserve performance on the original domain, require a fraction (typically 13\%, dependent on network architecture) of the number of parameters compared to standard fine-tuning procedures and converge in less cycles of training to a comparable or better level of performance. When coupled with standard network quantization techniques, we further reduce the parameter cost to around 3\% of the original with negligible or no loss in accuracy. The learned architecture can be controlled to switch between various learned representations, enabling a single network to solve a task from multiple different domains. We conduct extensive experiments showing the effectiveness of our method on a range of image classification tasks and explore different aspects of its behavior.

Amir Rasouli, John K. Tsotsos

In this paper, we investigate the effect of color space selection on detectability and discriminability of colored objects under various conditions. 20 color spaces from the literature are evaluated on a large dataset of simulated and real images. We measure the suitability of color spaces from two different perspectives: detectability and discriminability of various color groups. Through experimental evaluation, we found that there is no single optimal color space suitable for all color groups. The color spaces have different levels of sensitivity to different color groups and they are useful depending on the color of the sought object. Overall, the best results were achieved in both simulated and real images using color spaces C1C2C3, UVW and XYZ. In addition, using a simulated environment, we show a practical application of color space selection in the context of top-down control in active visual search. The results indicate that on average color space C1C2C3 followed by HSI and XYZ achieve the best time in searching for objects of various colors. Here, the right choice of color space can improve time of search on average by 20%. As part of our contribution, we also introduce a large dataset of simulated 3D objects