Peter Zeng, Weiling Li, Amie J. Paige, Zhengxiang Wang, Panagiotis Kaliosis, Dimitris Samaras, Gregory J. Zelinsky, Susan Brennan, Owen Rambow. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

A personalized model for scanpath prediction provides insights into the visual preferences and attention patterns of individual subjects. However, existing methods for training scanpath prediction models are data-intensive and cannot be effectively personalized to new individuals with only a few available examples. In this paper, we propose few-shot personalized scanpath prediction task (FS-PSP) and a novel method to address it, which aims to predict scan-paths for an unseen subject using minimal support data of that subject’s scanpath behavior. The key to our method’s adaptability is the Subject-Embedding Network (SE-Net), specifically designed to capture unique, individualized representations for each subject’s scanpaths. SE-Net generates subject embeddings that effectively distinguish between subjects while minimizing variability among scanpaths from the same individual. The personalized scanpath prediction model is then conditioned on these subject embeddings to produce accurate, personalized results. Experiments on multiple eye-tracking datasets demonstrate that our method excels in FS-PSP settings and does not require any fine-tuning steps at test time. Code is available at: https://github.com/cvlab-stonybrook/few-shot-scanpath

Most models of visual attention aim at predicting either top-down or bottom-up control, as studied using different visual search and free-viewing tasks. In this paper we propose the Human Attention Transformer (HAT), a single model that predicts both forms of attention control. HAT uses a novel transformer-based architecture and a simplified foveated retina that collectively create a spatio-temporal awareness akin to the dynamic visual working memory of humans. HAT not only establishes a new state-of-the-art in predicting the scanpath of fixations made during target-present and target-absent visual search and "taskless" free viewing, but also makes human gaze behavior interpretable. Unlike previous methods that rely on a coarse grid of fixation cells and experience information loss due to fixation discretization, HAT features a sequential dense prediction architecture and outputs a dense heatmap for each fixation, thus avoiding discretizing fixations. HAT sets a new standard in computational attention, which emphasizes effectiveness, generality, and interpretability. HAT's demonstrated scope and applicability will likely inspire the development of new attention models that can better predict human behavior in various attention-demanding scenarios. Code is available at https://github.com/cvlab-stonybrook/HAT.

Humans are extremely robust in our ability to perceive and recognize objects-we see faces in tea stains and can recognize friends on dark streets. Yet, neurocomputational models of primate object recognition have focused on the initial feed-forward pass of processing through the ventral stream and less on the top-down feedback that likely underlies robust object perception and recognition. Aligned with the generative approach, we propose that the visual system actively facilitates recognition by reconstructing the object hypothesized to be in the image. Top-down attention then uses this reconstruction as a template to bias feedforward processing to align with the most plausible object hypothesis. Building on auto-encoder neural networks, our model makes detailed hypotheses about the appearance and location of the candidate objects in the image by reconstructing a complete object representation from potentially incomplete visual input due to noise and occlusion. The model then leverages the best object reconstruction, measured by reconstruction error, to direct the bottom-up process of selectively routing low-level features, a top-down biasing that captures a core function of attention. We evaluated our model using the MNIST-C (handwritten digits under corruptions) and ImageNet-C (real-world objects under corruptions) datasets. Not only did our model achieve superior performance on these challenging tasks designed to approximate real-world noise and occlusion viewing conditions, but also better accounted for human behavioral reaction times and error patterns than a standard feedforward Convolutional Neural Network. Our model suggests that a complete understanding of object perception and recognition requires integrating top-down and attention feedback, which we propose is an object reconstruction.

Predicting human gaze is important in Human-Computer Interaction (HCI). However, to practically serve HCI applications, gaze prediction models must be scalable, fast, and accurate in their spatial and temporal gaze predictions. Recent scanpath prediction models focus on goaldirected attention (search). Such models are limited in their application due to a common approach relying on trained target detectors for all possible objects, and the availability of human gaze data for their training (both not scalable). In response, we pose a new task called ZeroGaze, a new variant of zero-shot learning where gaze is predicted for never-before-searched objects, and we develop a novel model, Gazeformer; to solve the ZeroGaze problem. In contrast to existing methods using object detector modules, Gazeformer encodes the target using a natural language model, thus leveraging semantic similarities in scanpath prediction. We use a transformer-based encoder-decoder architecture because transformers are particularly useful for generating contextual representations. Gazeformer surpasses other models by a large margin (19%-70%) on the ZeroGaze setting. It also outperforms existing target-detection models on standard gaze prediction for both target-present and target-absent search tasks. In addition to its improved performance, Gazeformer is more than five times faster than the state-of-the-art target-present visual search model. Code can be found at https://github.com/cvlab-stonybrook/Gazeformer/

We present a model for predicting visual attention during the free viewing of graphic design documents. While existing works on this topic have aimed at predicting static saliency of graphic designs, our work is the first attempt to predict both spatial attention and dynamic temporal order in which the document regions are fixated by gaze using a deep learning based model. We propose a two-stage model for predicting dynamic attention on such documents, with webpages being our primary choice of document design for demonstration. In the first stage, we predict the saliency maps for each of the document components (e.g. logos, banners, texts, etc. for webpages) conditioned on the type of document layout. These component saliency maps are then jointly used to predict the overall document saliency. In the second stage, we use these layout-specific component saliency maps as the state representation for an inverse reinforcement learning model of fixation scanpath prediction during document viewing. To test our model, we collected a new dataset consisting of eye movements from 41 people freely viewing 450 webpages (the largest dataset of its kind). Experimental results show that our model outperforms existing models in both saliency and scanpath prediction for webpages, and also generalizes very well to other graphic design documents such as comics, posters, mobile UIs, etc. and natural images.

Human gaze behavior prediction is important for behavioral vision and for computer vision applications. Most models mainly focus on predicting free-viewing behavior using saliency maps, but do not generalize to goal-directed behavior, such as when a person searches for a visual target object. We propose the first inverse reinforcement learning (IRL) model to learn the internal reward function and policy used by humans during visual search. We modeled the viewer's internal belief states as dynamic contextual belief maps of object locations. These maps were learned and then used to predict behavioral scanpaths for multiple target categories. To train and evaluate our IRL model we created COCO-Search18, which is now the largest dataset of high-quality search fixations in existence. COCO-Search18 has 10 participants searching for each of 18 target-object categories in 6202 images, making about 300,000 goal-directed fixations. When trained and evaluated on COCO-Search18, the IRL model outperformed baseline models in predicting search fixation scanpaths, both in terms of similarity to human search behavior and search efficiency. Finally, reward maps recovered by the IRL model reveal distinctive target-dependent patterns of object prioritization, which we interpret as a learned object context.

Observable reading behavior, the act of moving the eyes over lines of text, is highly stereotyped among the users of a language, and this has led to the development of reading detectors-methods that input windows of sequential fixations and output predictions of the fixation behavior during those windows being reading or skimming. The present study introduces a new method for reading detection using Region Ranking SVM (RRSVM). An SVM-based classifier learns the local oculomotor features that are important for real-time reading detection while it is optimizing for the global reading/skimming classification, making it unnecessary to hand-label local fixation windows for model training. This RRSVM reading detector was trained and evaluated using eye movement data collected in a laboratory context, where participants viewed modified web news articles and had to either read them carefully for comprehension or skim them quickly for the selection of keywords (separate groups). Ground truth labels were known at the global level (the instructed reading or skimming task), and obtained at the local level in a separate rating task. The RRSVM reading detector accurately predicted 82.5% of the global (article-level) reading/skimming behavior, with accuracy in predicting local window labels ranging from 72-95%, depending on how tuned the RRSVM was for local and global weights. With this RRSVM reading detector, a method now exists for near real-time reading detection without the need for hand-labeling of local fixation windows. With real-time reading detection capability comes the potential for applications ranging from education and training to intelligent interfaces that learn what a user is likely to know based on previous detection of their reading behavior.

Humans see only a tiny region at the center of their visual field with the highest visual acuity, a behavior known as foveation. Visual acuity reduces drastically towards the visual periphery. 'Foveated' video coding/compression techniques exploit this non-uniformity to gain significant efficiency by compressing more in the periphery and less in the center. We propose a practical and scalable method to use such a technique for video streaming service over the Internet. The essential idea is to use a commodity webcam on the user side to provide real-time gaze feedback to the server with the server sending appropriately coded video to the client player. We develop a multi-resolution video coding approach that is scalable in that it is possible to pre-code the video in a small number of copies for a given set of resolutions. The coding approach is designed to match the error performance of an eye tracker built using commodity webcams. We demonstrate that the technique is energy efficient and thus usable in mobile devices. We develop a methodology for performance evaluation of such a system when network budgets may vary and video quality may fluctuate. Finally, we present a comprehensive user study that demonstrates a bandwidth reduction of a factor of 2 for the same user satisfaction.

Learned region sparsity has achieved state-of-the-art performance in classification tasks by exploiting and integrating a sparse set of local information into global decisions. The underlying mechanism resembles how people sample information from an image with their eye movements when making similar decisions. In this paper we incorporate the biologically plausible mechanism of Inhibition of Return into the learned region sparsity model, thereby imposing diversity on the selected regions. We investigate how these mechanisms of sparsity and diversity relate to visual attention by testing our model on three different types of visual search tasks. We report state-of-the-art results in predicting the locations of human gaze fixations, even though our model is trained only on image-level labels without object location annotations. Notably, the classification performance of the extended model remains the same as the original. This work suggests a new computational perspective on visual attention mechanisms and shows how the inclusion of attention-based mechanisms can improve computer vision techniques.

Video segmentation is the task of grouping similar pixels in the spatio-temporal domain, and has become an important preprocessing step for subsequent video analysis. Most video segmentation and supervoxel methods output a hierarchy of segmentations, but while this provides useful multiscale information, it also adds difficulty in selecting the appropriate level for a task. In this work, we propose an efficient and robust video segmentation framework based on parametric graph partitioning (PGP), a fast, almost parameter free graph partitioning method that identifies and removes between-cluster edges to form node clusters. Apart from its computational efficiency, PGP performs clustering of the spatio-temporal volume without requiring a pre-specified cluster number or bandwidth parameters, thus making video segmentation more practical to use in applications. The PGP framework also allows processing sub-volumes, which further improves performance, contrary to other streaming video segmentation methods where sub-volume processing reduces performance. We evaluate the PGP method using the SegTrack v2 and Chen Xiph.org datasets, and show that it outperforms related state-of-the-art algorithms in 3D segmentation metrics and running time.

We posit that user behavior during natural viewing of images contains an abundance of information about the content of images as well as information related to user intent and user defined content importance. In this paper, we conduct experiments to better understand the relationship between images, the eye movements people make while viewing images, and how people construct natural language to describe images. We explore these relationships in the context of two commonly used computer vision datasets. We then further relate human cues with outputs of current visual recognition systems and demonstrate prototype applications for gaze-enabled detection and annotation.

Visual clutter, the perception of an image as being crowded and disordered, affects aspects of our lives ranging from object detection to aesthetics, yet relatively little effort has been made to model this important and ubiquitous percept. Our approach models clutter as the number of proto-objects segmented from an image, with proto-objects defined as groupings of superpixels that are similar in intensity, color, and gradient orientation features. We introduce a novel parametric method of merging superpixels by modeling mixture of Weibull distributions on similarity distance statistics, then taking the normalized number of proto-objects following partitioning as our estimate of clutter perception. We validated this model using a new $\text{90}-$image dataset of realistic scenes rank ordered by human raters for clutter, and showed that our method not only predicted clutter extremely well (Spearman's $\rho = 0.81$, $p < 0.05$), but also outperformed all existing clutter perception models and even a behavioral object segmentation ground truth. We conclude that the number of proto-objects in an image affects clutter perception more than the number of objects or features.

We propose a multi-resolution framework inspired by human visual search for general object detection. Different resolutions are represented using a coarse-to-fine feature hierarchy. During detection, the lower resolution features are initially used to reject the majority of negative windows at relatively low cost, leaving a relatively small number of windows to be processed in higher resolutions. This enables the use of computationally more expensive higher resolution features to achieve high detection accuracy. We applied this framework on Histograms of Oriented Gradient (HOG) features for object detection. Our multi-resolution detector produced better performance for pedestrian detection than state-of-the-art methods (Dalal and Triggs, 2005), and was faster during both training and testing. Testing our method on motorbikes and cars from the VOC database revealed similar improvements in both speed and accuracy, suggesting that our approach is suitable for realtime general object detection applications.

We combine local texture features (PCA-SIFT), global features (shape context), and spatial features within a single multi-layer AdaBoost model of object class recognition. The first layer selects PCA-SIFT and shape context features and combines the two feature types to form a strong classifier. Although previous approaches have used either feature type to train an AdaBoost model, our approach is the first to combine these complementary sources of information into a single feature pool and to use Adaboost to select those features most important for class recognition. The second layer adds to these local and global descriptions information about the spatial relationships between features. Through comparisons to the training sample, we first find the most prominent local features in Layer I, then capture the spatial relationships between these features in Layer 2. Rather than discarding this spatial information, we therefore use it to improve the strength of our classifier. We compared our method to (R. Fergus et al., 2003, A. Opelt et al., 2004, J. Thureson et al., 2004) and in all cases our approach outperformed these previous methods using a popular benchmark for object class recognition (R. Fergus et al., 2003). ROC equal error rates approached 99%. We also tested our method using a dataset of images that better equates the complexity between object and non-object images, and again found that our approach outperforms previous methods.

To investigate how top-down (TD) and bottom-up (BU) information is weighted in the guidance of human search behavior, we manipulated the proportions of BU and TD components in a saliency-based model. The model is biologically plausible and implements an artificial retina and a neuronal population code. The BU component is based on feature- contrast. The TD component is defined by a feature-template match to a stored target representation. We compared the model’s behavior at differ- ent mixtures of TD and BU components to the eye movement behavior of human observers performing the identical search task. We found that a purely TD model provides a much closer match to human behavior than any mixture model using BU information. Only when biological con- straints are removed (e.g., eliminating the retina) did a BU/TD mixture model begin to approximate human behavior.

We present a computational model of human eye movements in an ob- ject class detection task. The model combines state-of-the-art computer vision object class detection methods (SIFT features trained using Ad- aBoost) with a biologically plausible model of human eye movement to produce a sequence of simulated fixations, culminating with the acqui- sition of a target. We validated the model by comparing its behavior to the behavior of human observers performing the identical object class detection task (looking for a teddy bear among visually complex non- target objects). We found considerable agreement between the model and human data in multiple eye movement measures, including number of fixations, cumulative probability of fixating the target, and scanpath distance.