Computer Vision Tasks Research Articles

Tackling Heterogeneous Light Detection and Ranging-Camera Alignment Challenges in Dynamic Environments: A Review for Object Detection

Object detection is an essential computer vision task that identifies and locates objects within images or videos and is crucial for applications such as autonomous driving, robotics, and augmented reality. Light Detection and Ranging (LiDAR) and camera sensors are widely used for reliable object detection. These sensors produce heterogeneous data due to differences in data format, spatial resolution, and environmental responsiveness. Existing review articles on object detection predominantly focus on the statistical analysis of fusion algorithms, often overlooking the complexities of aligning data from these distinct modalities, especially dynamic environment data alignment. This paper addresses the challenges of heterogeneous LiDAR-camera alignment in dynamic environments by surveying over 20 alignment methods for three-dimensional (3D) object detection, focusing on research published between 2019 and 2024. This study introduces the core concepts of multimodal 3D object detection, emphasizing the importance of integrating data from different sensor modalities for accurate object recognition in dynamic environments. The survey then delves into a detailed comparison of recent heterogeneous alignment methods, analyzing critical approaches found in the literature, and identifying their strengths and limitations. A classification of methods for aligning heterogeneous data in 3D object detection is presented. This paper also highlights the critical challenges in aligning multimodal data, including dynamic environments, sensor fusion, scalability, and real-time processing. These limitations are thoroughly discussed, and potential future research directions are proposed to address current gaps and advance the state-of-the-art. By summarizing the latest advancements and highlighting open challenges, this survey aims to stimulate further research and innovation in heterogeneous alignment methods for multimodal 3D object detection, thereby pushing the boundaries of what is currently achievable in this rapidly evolving domain.

Swin Wavelet Transformer (SWT): Mixing Tokens with Wavelet and Multiwavelet Transforms

The Swin Transformer possess a hierarchical structure, a robust structure, and an efficient self-attention mechanism. This makes it superior in performance for a wide variety of artificial intelligence and machine training tasks. It has revolutionized the field of digital signal processing and its applications by creating vision in multiple fields. It uses non-overlapping windows which leads to creating a distinct perspective and understanding of the context of the digital signals. However, it possess a very complicated hierarchical structure an complexities together. By reducing calculations and complexities together. This will lead to robust performance that make it a powerful tool for a wide range of computer vision tasks. In this research, simplified symbol mixing methods were developed for coding structures similar to transformers, by forming various semantics in the text through linear mixing transformation and in combination with nonlinearity in the feed-forward layers. Hence, we were able to successfully propose a Swin Wavelet Transformer (SWT) model in which the self-attention sublayer was replaced by a Wavelet transform. In a second attempt, the wavelet transform was replaced by the multi-wavelet transform. These two proposed models are achieved a better performance from their FNets and BERT counterparts and are highly competitive with traditional and efficient transformers.

Enhancing low-illumination imagery using a Deep Convolutional Generative Adversarial Network with weight regularization (DCGAN-WR) and Zero-Reference Deep Curve Estimation (DCE)

Improving low-light images to enhance prediction in various applications has a greater advantage. A two-pronged approach that employs Deep Convolutional Generative Adversarial Networks with weight regularization (DCGAN-WR) and Zero-Reference Deep Curve Estimation (DCE) was used. The model was trained using the LOL dataset, and the results showed significant improvements in image quality. The DCGAN is fine-tuned with Group Lasso regularization to enhance the performance. The DCGAN-WR model is shown to enhance images realistically, demonstrating its capacity to learn characteristics and texture representations from low-light input. Empirical and simulated image comparisons demonstrate remarkable performance under demanding low-light settings. Moreover, the DCE model employs a novel approach that considers color constancy loss, light smoothness, and spatial consistency. Information about the learning dynamics and curve parameter changing capabilities of the model can be visualized by loss function graphs, which aim to maximize picture quality. Compared to the original images, Images generated by the DCE models maintain color accuracy, increase exposure levels, and preserve spatial coherence. A solution for low-illumination image enhancement is achieved through the proposed model DCGAN-WR and DCE. Genuine details are recorded by the GAN model, while the DCE adjusts the exposure levels and color balance to produce improved, aesthetically pleasing, and contextually accurate images. The proposed approach not only outperforms the other methods on the LOL dataset but also exhibits potential for practical use in computer vision tasks, which require higher image quality for precise analysis and interpretation.

Weakly-supervised Semantic Segmentation with Image-level Labels: From Traditional Models to Foundation Models

The rapid development of deep learning has driven significant progress in image semantic segmentation—a fundamental task in computer vision. Semantic segmentation algorithms often depend on the availability of pixel-level labels (i.e., masks of objects), which are expensive, time-consuming, and labor-intensive. Weakly-supervised semantic segmentation (WSSS) is an effective solution to avoid such labeling. It utilizes only partial or incomplete annotations and provides a cost-effective alternative to fully-supervised semantic segmentation. In this journal, our focus is on the WSSS with image-level labels, which is the most challenging form of WSSS. Our work has two parts. First, we conduct a comprehensive survey on traditional methods, primarily focusing on those presented at premier research conferences. We categorize them into four groups based on where their methods operate: pixel-wise, image-wise, cross-image, and external data. Second, we investigate the applicability of visual foundation models, such as the Segment Anything Model (SAM), in the context of WSSS. We scrutinize SAM in two intriguing scenarios: text prompting and zero-shot learning. We provide insights into the potential and challenges of deploying visual foundational models for WSSS, facilitating future developments in this exciting research area. Our code is provided at this link.

IMMEDIATE OBJECT SPOTTING

Object Spotting has become one of the most essential and challenging tasks in computer vision, particularly in the context of deep learning advancements. This paper reviews the evolution of object detection techniques, highlighting how deep neural networks have transformed the field. We classify the state-of-the-art detection algorithms into three main categories: anchor-based, anchor-free, and transformer-based detectors, each with distinct methodologies for identifying objects in images. This survey includes a comparison of key convolutional neural network (CNN) architectures used for object detection, evaluating their speed-accuracy trade-offs, quality metrics, and training strategies. We discuss the strengths and limitations of these models and offer insights into the future directions of research in object spotting. In parallel, immediate object spotting using OpenCV has gained prominence for its ability to identify and localize objects within live video streams, particularly by leveraging pre-trained models like YOLOv3. This approach processes each video frame in real time, with preprocessing steps such as resizing and normalization to ensure compatibility with the deep learning model. YOLOv3’s efficiency enables it to detect and classify objects quickly, drawing bounding boxes around detected objects and displaying them in real time. This project highlights how combining OpenCV with deep learning models can achieve effective, scalable object spotting while ensuring detection accuracy. The paper also explores the challenges and potential future advancements in improving detection accuracy, robustness, and compatibility with modern hardware systems for diverse applications. Keywords:Immediate object spotting, OpenCV, YOLOv3, Deep learning models, Live video stream, Pre- processing (resizing, normalization), Object identification and localization, Bounding boxes, Detection accuracy, System compatibility

SiamRhic: Improved Cross-Correlation and Ranking Head-Based Siamese Network for Object Tracking in Remote Sensing Videos

Object tracking in remote sensing videos is a challenging task in computer vision. Recent advances in deep learning have sparked significant interest in tracking algorithms based on Siamese neural networks. However, many existing algorithms fail to deliver satisfactory performance in complex scenarios due to challenging conditions and limited computational resources. Thus, enhancing tracking efficiency and improving algorithm responsiveness in complex scenarios are crucial. To address tracking drift caused by similar objects and background interference in remote sensing image tracking, we propose an enhanced Siamese network based on the SiamRhic architecture, incorporating a cross-correlation and ranking head for improved object tracking. We first use convolutional neural networks for feature extraction and integrate the CBAM (Convolutional Block Attention Module) to enhance the tracker’s representational capacity, allowing it to focus more effectively on the objects. Additionally, we replace the original depth-wise cross-correlation operation with asymmetric convolution, enhancing both speed and performance. We also introduce a ranking loss to reduce the classification confidence of interference objects, addressing the mismatch between classification and regression. We validate the proposed algorithm through experiments on the OTB100, UAV123, and OOTB remote sensing datasets. Specifically, SiamRhic achieves success, normalized precision, and precision rates of 0.533, 0.786, and 0.812, respectively, on the OOTB benchmark. The OTB100 benchmark achieves a success rate of 0.670 and a precision rate of 0.892. Similarly, in the UAV123 benchmark, SiamRhic achieves a success rate of 0.621 and a precision rate of 0.823. These results demonstrate the algorithm’s high precision and success rates, highlighting its practical value.

Selecting the Right Metric: A Detailed Study on Image Segmentation Evaluation

<p dir="ltr"><span>Image segmentation is critical role in various computer vision tasks, such as object recognition, medical imaging, autonomous driving, and document analysis. To evaluate the performance of segmentation algorithms, various metrics have been developed, each providing insights into different aspects of segmentation accuracy, object completeness, and boundary precision. This paper presents a comprehensive review of the key evaluation metrics used in image segmentation, including popular metrics such as Intersection over Union, Dice Coefficient, F1-score, and boundary-based metrics like average and maximum boundary distance. We explore the taxonomy of evaluation techniques, distinguishing between supervised and unsupervised methods, and discuss the strengths, limitations, and applications of each metric. Furthermore, we address the challenges associated with image segmentation, including handling noise, occlusions, illumination variation, and the scarcity of annotated datasets for training. By offering a detailed analysis of evaluation metrics, this paper aims to guide researchers and practitioners in selecting appropriate metrics for specific tasks and improving algorithm performance through informed comparisons and parameter tuning.</span></p><div><span><br /></span></div>

Computer vision methods for side scan sonar imagery

Abstract Side-scan sonar data presents a significant challenge to analysts due to the size of the material that requires processing. Automatic recognition systems can help in the interpretation and filtration of the data, reducing the time needed for analysis. Deep learning architectures have been employed for computer vision tasks in the underwater domain. The objective of this work is to review current deep learning methods for automatic image classification, object detection, semantic segmentation, and instance segmentation through statistical analysis of published studies. Further, this work includes an inventory of publicly available side-scan sonar data followed by a discussion on augmentation methods applied to side-scan sonar imagery. The study discusses current challenges in the automatic underwater target detection field, highlights main gaps within the domain and identifies potential research directions. Our proposal includes the creation of a standardised benchmark and the use of a common set of metrics allowing for consistent evaluation of future research works.

Multi-scale Perception Enhancement of Structural Patch Decomposition and Fusion for low light imaging

Low light imaging usually exhibits defects such as low contrast, low signal-to-noise ratio, low image quality and severe color distortion. Enhancing images taken in such scenes is efficient and rewarding for image restoration and computer vision tasks under demanding imaging conditions. Existing low light image enhancement(LLIE) algorithms rarely take the consistency of image into account, showing insufficiently realistic restoration results. To solve these problem, we propose a LLIE method based on Multi-scale Perception Enhancement of Structural Patch Decomposition and Fusion (MPESPDF-LLIE) to restore images captured in a low light scene. In this model, each image patch is firstly decomposed into four components: perception gain, average intensity, signal strength and signal structure. Next the enhancement model produces the global perception gain and the original enhanced image patches by the inverted dehazing strategy. Based on the original enhanced image patches, an efficient way is used to enhance the signal strength and signal structure in patches. Then a brightness compensation is adopted to improve average intensity. Finally the four components are fused and aggregated to the final image in a multi-scale framework. Experimental results demonstrate that MPESPDF-LLIE is advantageous and effective in preserving detail information and maintaining image consistency, thus MPESPDF-LLIE compensates for the shortcomings of low-light imaging and improves image quality. The source code is available at: https://github.com/LYor61/MPESPDF-LLIE.

DESRGAN: Detail-enhanced generative adversarial networks for small sample single image super-resolution

Image super-resolution involves reconstructing a blurry, low-resolution image with limited information into a clear, high-resolution image containing more detailed information. The images generated by super-resolution reconstruction can enhance the performance of downstream computer vision tasks, and hold wide application prospects in fields such as industrial fault detection, plant phenotype parameter extraction, medical imaging, and more. High-frequency components in images, such as edges and texture details, typically require more attention. However, when the training samples are limited, effectively recovering clear high-frequency details of images becomes highly challenging. Therefore, this paper proposes a single-image super-resolution method based on generative adversarial networks, named DESRGAN. Compared to existing methods, DESRGAN achieves better reconstruction of image details even with a limited number of training samples. DESRGAN introduces several key innovations: a shallow generator structure to address overfitting issues in small sample scenarios, a dual-stream feature extraction network with dilated convolutions to capture multi-scale contextual information and expand the receptive field, and an artifact loss designed to eliminate artifacts and preserve the true high-frequency details of the super-resolved images. Extensive ablation experiments and comparative studies with multiple state-of-the-art models are conducted on two small sample datasets, "Root" and "Leaves," as well as five publicly available datasets. The results demonstrate that the proposed DESRGAN achieves superior performance in small sample single image super-resolution tasks, with improvements of 1.39 dB in PSNR and 0.013 in SSIM. The generated high-resolution images exhibit clear texture and edge structures, presenting favorable subjective visual effects. Moreover, the model displays strong generalization capabilities.

Improved Osteoporosis Prediction in Breast Cancer Patients Using a Novel Semi-Foundational Model.

Small cohorts of certain disease states are common especially in medical imaging. Despite the growing culture of data sharing, information safety often precludes open sharing of these datasets for creating generalizable machine learning models. To overcome this barrier and maintain proper health information protection, foundational models are rapidly evolving to provide deep learning solutions that have been pretrained on the native feature spaces of the data. Although this has been optimized in Large Language Models (LLMs), there is still a sparsity of foundational models for computer vision tasks. It is in this space that we provide an investigation into pretraining Visual Geometry Group (VGG)-16, Residual Network (ResNet)-50, and Dense Network (DenseNet)-121 on an unrelated dataset of 8500 chest CTs which was subsequently fine-tuned to classify bone mineral density (BMD) in 199 breast cancer patients using the L1 vertebra on CT. These semi-foundational models showed significant improved ternary classification into mild, moderate, and severe demineralization in comparison to ground truth Hounsfield Unit (HU) measurements in trabecular bone with the semi-foundational ResNet50 architecture demonstrating the best relative performance. Specifically, the holdout testing AUC was 0.99 (p-value < 0.05, ANOVA versus no pretraining versus ImageNet transfer learning) and F1-score 0.99 (p-value < 0.05) for the holdout testing set. In this study, the use of a semi-foundational model trained on the native feature space of CT provided improved classification in a completely disparate disease state with different window levels. Future implementation with these models may provide better generalization despite smaller numbers of a disease state to be classified.

GAFAR revisited—Exploring the limits of point cloud registration on sparse subsets

Robust estimation of a rigid transformation aligning two point clouds is a fundamental building block in many tasks in computer vision and robotics, such as mapping, (re-)localization, SLAM or 3D scanning, and many more. Deep-Learning based registration methods have proven superior in terms of robustness to outliers and bad initializations, yet their resource needs often render them unsuitable for applications where space, energy and real-time constraints come into play. Since - provided a good initialization - conventional registration algorithms like ICP prove to be fast and accurate, we may relax the requirements of absolute precision of a learning-based matcher and instead focus our attention on efficiency, robustness and generalization.In prior work, we introduced our small, fast and light-weight registration algorithm GAFAR, which works on sparse subsets of point clouds and exhibits a small enough footprint to be deployed on edge devices in mobile applications, while still providing accurate results and promising generalization ability. Based thereon, we develop this idea further towards applicability on real world data with improvements to the algorithm as well as further evaluations and analyses. We showcase this by applying it to Kitti Odometry Benchmark and 3DMatch dataset as well as demonstrating its usability on edge devices. The code and trained weights are published in https://github.com/mordecaimalignatius/GAFAR/.

Information gap based knowledge distillation for occluded facial expression recognition

Facial Expression Recognition (FER) with occlusion presents a challenging task in computer vision because facial occlusions result in poor visual data features. Recently, the region attention technique has been introduced to address this problem by researchers, which make the model perceive occluded regions of the face and prioritize the most discriminative non-occluded regions. However, in real-world scenarios, facial images are influenced by various factors, including hair, masks and sunglasses, making it difficult to extract high-quality features from these occluded facial images. This inevitably limits the effectiveness of attention mechanisms. In this paper, we observe a correlation in facial emotion features from the same image, both with and without occlusion. This correlation contributes to addressing the issue of facial occlusions. To this end, we propose a Information Gap based Knowledge Distillation (IGKD) to explore the latent relationship. Specifically, our approach involves feeding non-occluded and masked images into separate teacher and student networks. Due to the incomplete emotion information in the masked images, there exists an information gap between the teacher and student networks. During training, we aim to minimize this gap to enable the student network to learn this relationship. To enhance the teacher’s guidance, we introduce a joint learning strategy where the teacher conducts knowledge distillation on the student during the training of the teacher. Additionally, we introduce two novel constraints, called knowledge learn and knowledge feedback loss, to supervise and optimize both the teacher and student networks. The reported experimental results show that IGKD outperforms other algorithms on four benchmark datasets. Specifically, our IGKD achieves 87.57% on Occlusion-RAF-DB, 87.33% on Occlusion-FERPlus, 64.86% on Occlusion-AffectNet, and 73.25% on FED-RO, clearly demonstrating its effectiveness and robustness.

Advances in Deep Reinforcement Learning for Computer Vision Applications

Deep Reinforcement Learning (DRL) has become very popular for computer vision (CV), solving mostly visually complex environments with decision making and dynamic adaption to different situations. This article provides an introduction to the basic concepts of deep reinforcement learning with a special focus on their applications in computer vision tasks, including challenging problems and emerging solutions. It reviews various DRL algorithms such as Q-learning, policy gradient methods, and Actor-Critic models, explaining much of the modifications that are done to make it work on high-dimensional visual problems. Different applications of DRL in major CV applications like object detection, image segmentation, target tracking and image generation are reviewed to demonstrate the power as well as limitations of DRL in practice. More importantly, the novel paradigms such as hierarchical policy learning, adaptive reward design multi-task reinforcement learning and domain adaptation can be viewed as promising premises to improve model efficiency and generalizability in multiple scenarios. Finally, this paper mentions a few of the existing challenges like computational power cost and sample efficiency; as well as future paths for enhancing that can widen devout reinforcement learning in computer vision. Through this comprehensive overview, we aim to shed light on the promising synergies between DRL and CV, while identifying key areas for future research and application.

DeepCervix: enhancing cervical cancer detection through transfer learning with VGG-16 architecture

Cervical cancer remains a significant global health concern, emphasizing the urgent need for improved detection methods to ensure timely treatment. This research introduces a sophisticated methodology leveraging recent advances in medical imaging and deep learning algorithms to enhance the accuracy and efficiency of cervical cancer detection. The proposed approach comprises meticulous data preprocessing to ensure the integrity of input images, followed by the training of deep learning models including ResNet-50, AlexNet, and VGG-16, renowned for their performance in computer vision tasks. Evaluation metrics such as accuracy, precision, recall, and F1-score demonstrate the efficacy of the methodology, with an outstanding accuracy rate of 98% achieved. The model’s proficiency in accurately distinguishing healthy cervical tissue from cancerous tissue is underscored by precision, recall, and F1-score values. The primary strength of this deep learning-based approach lies in its potential for early detection, promising significant impact on cervical cancer diagnosis and treatment outcomes. This methodology contributes to advancements in medical imaging techniques, facilitating improved outcomes in cervical cancer detection and treatment.

Fast Learning of Signed Distance Functions From Noisy Point Clouds via Noise to Noise Mapping.

Learning signed distance functions (SDFs) from point clouds is an important task in 3D computer vision. However, without ground truth signed distances, point normals or clean point clouds, current methods still struggle from learning SDFs from noisy point clouds. To overcome this challenge, we propose to learn SDFs via a noise to noise mapping, which does not require any clean point cloud or ground truth supervision. Our novelty lies in the noise to noise mapping which can infer a highly accurate SDF of a single object or scene from its multiple or even single noisy observations. We achieve this by a novel loss which enables statistical reasoning on point clouds and maintains geometric consistency although point clouds are irregular, unordered and have no point correspondence among noisy observations. To accelerate training, we use multi-resolution hash encodings implemented in CUDA in our framework, which reduces our training time by a factor of ten, achieving convergence within one minute. We further introduce a novel schema to improve multi-view reconstruction by estimating SDFs as a prior. Our evaluations under widely-used benchmarks demonstrate our superiority over the state-of-the-art methods in surface reconstruction from point clouds or multi-view images, point cloud denoising and upsampling.

CAP-UDF: Learning Unsigned Distance Functions Progressively From Raw Point Clouds With Consistency-Aware Field Optimization.

Surface reconstruction for point clouds is an important task in 3D computer vision. Most of the latest methods resolve this problem by learning signed distance functions from point clouds, which are limited to reconstructing closed surfaces. Some other methods tried to represent open surfaces using unsigned distance functions (UDF) which are learned from ground truth distances. However, the learned UDF is hard to provide smooth distance fields due to the discontinuous character of point clouds. In this paper, we propose CAP-UDF, a novel method to learn consistency-aware UDF from raw point clouds. We achieve this by learning to move queries onto the surface with a field consistency constraint, where we also enable to progressively estimate a more accurate surface. Specifically, we train a neural network to gradually infer the relationship between queries and the approximated surface by searching for the moving target of queries in a dynamic way. Meanwhile, we introduce a polygonization algorithm to extract surfaces using the gradients of the learned UDF. We conduct comprehensive experiments in surface reconstruction for point clouds, real scans or depth maps, and further explore our performance in unsupervised point normal estimation, which demonstrate non-trivial improvements of CAP-UDF over the state-of-the-art methods.

Secrets of Event-Based Optical Flow, Depth and Ego-Motion Estimation by Contrast Maximization.

Event cameras respond to scene dynamics and provide signals naturally suitable for motion estimation with advantages, such as high dynamic range. The emerging field of event-based vision motivates a revisit of fundamental computer vision tasks related to motion, such as optical flow and depth estimation. However, state-of-the-art event-based optical flow methods tend to originate in frame-based deep-learning methods, which require several adaptations (data conversion, loss function, etc.) as they have very different properties. We develop a principled method to extend the Contrast Maximization framework to estimate dense optical flow, depth, and ego-motion from events alone. The proposed method sensibly models the space-time properties of event data and tackles the event alignment problem. It designs the objective function to prevent overfitting, deals better with occlusions, and improves convergence using a multi-scale approach. With these key elements, our method ranks first among unsupervised methods on the MVSEC benchmark and is competitive on the DSEC benchmark. Moreover, it allows us to simultaneously estimate dense depth and ego-motion, exposes the limitations of current flow benchmarks, and produces remarkable results when it is transferred to unsupervised learning settings. Along with various downstream applications shown, we hope the proposed method becomes a cornerstone on event-based motion-related tasks.

Optical Flow as Spatial-Temporal Attention Learners.

Optical flow is an indispensable building block for various important computer vision tasks, including motion estimation, object tracking, and disparity measurement. To date, the dominant methods are CNN-based, leaving plenty of room for improvement. In this work, we propose TransFlow, a transformer architecture for optical flow estimation. Compared to dominant CNN-based methods, TransFlow demonstrates three advantages. First, it provides more accurate correlation and trustworthy matching in flow estimation by utilizing spatial self-attention and cross-attention mechanisms between adjacent frames to effectively capture global dependencies; Second, it recovers more compromised information (e.g., occlusion and motion blur) in flow estimation through long-range temporal association in dynamic scenes; Third, it introduces a concise self-learning paradigm, eliminating the need for complex and laborious multi-stage pre-training procedures. The versatility and superiority of TransFlow extend seamlessly to 3D scene motion, yielding competitive outcomes in 3D scene flow estimation. Our approach attains state-of-the-art results on benchmark datasets such as Sintel and KITTI-15, while also exhibiting exceptional performance on downstream tasks, including video object detection using the ImageNet VID dataset, video frame interpolation using the GoPro dataset, and video stabilization using the DeepStab dataset. We believe that the effectiveness of TransFlow positions it as a flexible baseline for both optical flow and scene flow estimation, offering promising avenues for future research and development.

Learning From Human Attention for Attribute-Assisted Visual Recognition.

With prior knowledge of seen objects, humans have a remarkable ability to recognize novel objects using shared and distinct local attributes. This is significant for the challenging tasks of zero-shot learning (ZSL) and fine-grained visual classification (FGVC), where the discriminative attributes of objects have played an important role. Inspired by human visual attention, neural networks have widely exploited the attention mechanism to learn the locally discriminative attributes for challenging tasks. Though greatly promoted the development of these fields, existing works mainly focus on learning the region embeddings of different attribute features and neglect the importance of discriminative attribute localization. It is also unclear whether the learned attention truly matches the real human attention. To tackle this problem, this paper proposes to employ real human gaze data for visual recognition networks to learn from human attention. Specifically, we design a unified Attribute Attention Network (A 2Net) that learns from human attention for both ZSL and FGVC tasks. The overall model consists of an attribute attention branch and a baseline classification network. On top of the image feature maps provided by the baseline classification network, the attribute attention branch employs attribute prototypes to produce attribute attention maps and attribute features. The attribute attention maps are converted to gaze-like attentions to be aligned with real human gaze attention. To guarantee the effectiveness of attribute feature learning, we further align the extracted attribute features with attribute-defined class embeddings. To facilitate learning from human gaze attention for the visual recognition problems, we design a bird classification game to collect real human gaze data using the CUB dataset via an eye-tracker device. Experiments on ZSL and FGVC tasks without/with real human gaze data validate the benefits and accuracy of our proposed model. This work supports the promising benefits of collecting human gaze datasets and automatic gaze estimation algorithms learning from human attention for high-level computer vision tasks.