Ali Borji

Ali Borji

For nearly a decade, the COCO dataset has been the central test bed of research in object detection. According to the recent benchmarks, however, it seems that performance on this dataset has started to saturate. One possible reason can be that perhaps it is not large enough for training deep models. To address this limitation, here we introduce two complementary datasets to COCO: i) COCO_OI, composed of images from COCO and OpenImages (from their 80 classes in common) with 1,418,978 training bounding boxes over 380,111 images, and 41,893 validation bounding boxes over 18,299 images, and ii) ObjectNet_D containing objects in daily life situations (originally created for object recognition known as ObjectNet; 29 categories in common with COCO). The latter can be used to test the generalization ability of object detectors. We evaluate some models on these datasets and pinpoint the source of errors. We encourage the community to utilize these datasets for training and testing object detection models. Code and data is available at https://github.com/aliborji/COCO_OI.

Ali Borji

Large language models have been demonstrated to be valuable in different fields. ChatGPT, developed by OpenAI, has been trained using massive amounts of data and simulates human conversation by comprehending context and generating appropriate responses. It has garnered significant attention due to its ability to effectively answer a broad range of human inquiries, with fluent and comprehensive answers surpassing prior public chatbots in both security and usefulness. However, a comprehensive analysis of ChatGPT's failures is lacking, which is the focus of this study. Eleven categories of failures, including reasoning, factual errors, math, coding, and bias, are presented and discussed. The risks, limitations, and societal implications of ChatGPT are also highlighted. The goal of this study is to assist researchers and developers in enhancing future language models and chatbots.

Ali Borji, Sikun Lin

We present SplitMixer, a simple and lightweight isotropic MLP-like architecture, for visual recognition. It contains two types of interleaving convolutional operations to mix information across spatial locations (spatial mixing) and channels (channel mixing). The first one includes sequentially applying two depthwise 1D kernels, instead of a 2D kernel, to mix spatial information. The second one is splitting the channels into overlapping or non-overlapping segments, with or without shared parameters, and applying our proposed channel mixing approaches or 3D convolution to mix channel information. Depending on design choices, a number of SplitMixer variants can be constructed to balance accuracy, the number of parameters, and speed. We show, both theoretically and experimentally, that SplitMixer performs on par with the state-of-the-art MLP-like models while having a significantly lower number of parameters and FLOPS. For example, without strong data augmentation and optimization, SplitMixer achieves around 94% accuracy on CIFAR-10 with only 0.28M parameters, while ConvMixer achieves the same accuracy with about 0.6M parameters. The well-known MLP-Mixer achieves 85.45% with 17.1M parameters. On CIFAR-100 dataset, SplitMixer achieves around 73% accuracy, on par with ConvMixer, but with about 52% fewer parameters and FLOPS. We hope that our results spark further research towards finding more efficient vision architectures and facilitate the development of MLP-like models. Code is available at https://github.com/aliborji/splitmixer.

Ali Borji

Hybrid images is a technique to generate images with two interpretations that change as a function of viewing distance. It has been utilized to study multiscale processing of images by the human visual system. Using 63,000 hybrid images across 10 fruit categories, here we show that predictions of deep learning vision models qualitatively matches with the human perception of these images. Our results provide yet another evidence in support of the hypothesis that Convolutional Neural Networks (CNNs) and Transformers are good at modeling the feedforward sweep of information in the ventral stream of visual cortex. Code and data is available at https://github.com/aliborji/hybrid_images.git.

Ali Borji

Almost all adversarial attacks are formulated to add an imperceptible perturbation to an image in order to fool a model. Here, we consider the opposite which is adversarial examples that can fool a human but not a model. A large enough and perceptible perturbation is added to an image such that a model maintains its original decision, whereas a human will most likely make a mistake if forced to decide (or opt not to decide at all). Existing targeted attacks can be reformulated to synthesize such adversarial examples. Our proposed attack, dubbed NKE, is similar in essence to the fooling images, but is more efficient since it uses gradient descent instead of evolutionary algorithms. It also offers a new and unified perspective into the problem of adversarial vulnerability. Experimental results over MNIST and CIFAR-10 datasets show that our attack is quite efficient in fooling deep neural networks. Code is available at https://github.com/aliborji/NKE.

Ali Borji

The field of image synthesis has made great strides in the last couple of years. Recent models are capable of generating images with astonishing quality. Fine-grained evaluation of these models on some interesting categories such as faces is still missing. Here, we conduct a quantitative comparison of three popular systems including Stable Diffusion, Midjourney, and DALL-E 2 in their ability to generate photorealistic faces in the wild. We find that Stable Diffusion generates better faces than the other systems, according to the FID score. We also introduce a dataset of generated faces in the wild dubbed GFW, including a total of 15,076 faces. Furthermore, we hope that our study spurs follow-up research in assessing the generative models and improving them. Data and code are available at data and code, respectively.

Ali Borji

The primary aim of this manuscript is to underscore a significant limitation in current deep learning models, particularly vision models. Unlike human vision, which efficiently selects only the essential visual areas for further processing, leading to high speed and low energy consumption, deep vision models process the entire image. In this work, we examine this issue from a broader perspective and propose two solutions that could pave the way for the next generation of more efficient vision models. In the first solution, convolution and pooling operations are selectively applied to altered regions, with a change map sent to subsequent layers. This map indicates which computations need to be repeated. In the second solution, only the modified regions are processed by a semantic segmentation model, and the resulting segments are inserted into the corresponding areas of the previous output map. The code is available at https://github.com/aliborji/spatial_attention.

Ali Borji

We show that it is possible to predict which deep network has generated a given logit vector with accuracy well above chance. We utilize a number of networks on a dataset, initialized with random weights or pretrained weights, as well as fine-tuned networks. A classifier is then trained on the logit vectors of the trained set of this dataset to map the logit vector to the network index that has generated it. The classifier is then evaluated on the test set of the dataset. Results are better with randomly initialized networks, but also generalize to pretrained networks as well as fine-tuned ones. Classification accuracy is higher using unnormalized logits than normalized ones. We find that there is little transfer when applying a classifier to the same networks but with different sets of weights. In addition to help better understand deep networks and the way they encode uncertainty, we anticipate our finding to be useful in some applications (e.g. tailoring an adversarial attack for a certain type of network). Code is available at https://github.com/aliborji/logits.

Ali Borji

Object detection is a fundamental vision task. It has been highly researched in academia and has been widely adopted in industry. Average Precision (AP) is the standard score for evaluating object detectors. Our understanding of the subtleties of this score, however, is limited. Here, we quantify the sensitivity of AP to bounding box perturbations and show that AP is very sensitive to small translations. Only one pixel shift is enough to drop the mAP of a model by 8.4%. The mAP drop over small objects with only one pixel shift is 23.1%. The corresponding numbers when ground-truth (GT) boxes are used as predictions are 23% and 41.7%, respectively. These results explain why achieving higher mAP becomes increasingly harder as models get better. We also investigate the effect of box scaling on AP. Code and data is available at https://github.com/aliborji/AP_Box_Perturbation.

Ali Borji

The ability of image and video generation models to create photorealistic images has reached unprecedented heights, making it difficult to distinguish between real and fake images in many cases. However, despite this progress, a gap remains between the quality of generated images and those found in the real world. To address this, we have reviewed a vast body of literature from both academic publications and social media to identify qualitative shortcomings in image generation models, which we have classified into five categories. By understanding these failures, we can identify areas where these models need improvement, as well as develop strategies for detecting deep fakes. The prevalence of deep fakes in today's society is a serious concern, and our findings can help mitigate their negative impact.

Andrey Moskalenko, Alexey Bryncev, Dmitry Vatolin et al.

This paper reviews the Challenge on Video Saliency Prediction at AIM 2024. The goal of the participants was to develop a method for predicting accurate saliency maps for the provided set of video sequences. Saliency maps are widely exploited in various applications, including video compression, quality assessment, visual perception studies, the advertising industry, etc. For this competition, a previously unused large-scale audio-visual mouse saliency (AViMoS) dataset of 1500 videos with more than 70 observers per video was collected using crowdsourced mouse tracking. The dataset collection methodology has been validated using conventional eye-tracking data and has shown high consistency. Over 30 teams registered in the challenge, and there are 7 teams that submitted the results in the final phase. The final phase solutions were tested and ranked by commonly used quality metrics on a private test subset. The results of this evaluation and the descriptions of the solutions are presented in this report. All data, including the private test subset, is made publicly available on the challenge homepage - https://challenges.videoprocessing.ai/challenges/video-saliency-prediction.html.

Ali Borji

My goal in this paper is twofold: to study how well deep models can understand the images generated by DALL-E 2 and Midjourney, and to quantitatively evaluate these generative models. Two sets of generated images are collected for object recognition and visual question answering (VQA) tasks. On object recognition, the best model, out of 10 state-of-the-art object recognition models, achieves about 60\% and 80\% top-1 and top-5 accuracy, respectively. These numbers are much lower than the best accuracy on the ImageNet dataset (91\% and 99\%). On VQA, the OFA model scores 77.3\% on answering 241 binary questions across 50 images. This model scores 94.7\% on the binary VQA-v2 dataset. Humans are able to recognize the generated images and answer questions on them easily. We conclude that a) deep models struggle to understand the generated content, and may do better after fine-tuning, and b) there is a large distribution shift between the generated images and the real photographs. The distribution shift appears to be category-dependent. Data is available at: https://drive.google.com/file/d/1n2nCiaXtYJRRF2R73-LNE3zggeU_HeH0/view?usp=sharing.

Ali Borji

We introduce a new test set for visual question answering (VQA) called BinaryVQA to push the limits of VQA models. Our dataset includes 7,800 questions across 1,024 images and covers a wide variety of objects, topics, and concepts. For easy model evaluation, we only consider binary questions. Questions and answers are formulated and verified carefully and manually. Around 63% of the questions have positive answers. The median number of questions per image and question length are 7 and 5, respectively. The state of the art OFA model achieves 75% accuracy on BinaryVQA dataset, which is significantly lower than its performance on the VQA v2 test-dev dataset (94.7%). We also analyze the model behavior along several dimensions including: a) performance over different categories such as text, counting and gaze direction, b) model interpretability, c) the effect of question length on accuracy, d) bias of models towards positive answers and introduction of a new score called the ShuffleAcc, and e) sensitivity to spelling and grammar errors. Our investigation demonstrates the difficulty of our dataset and shows that it can challenge VQA models for next few years. Data and code are publicly available at: DATA and CODE.

Ali Borji

Short answer: Yes, Long answer: No! Indeed, research on adversarial robustness has led to invaluable insights helping us understand and explore different aspects of the problem. Many attacks and defenses have been proposed over the last couple of years. The problem, however, remains largely unsolved and poorly understood. Here, I argue that the current formulation of the problem serves short term goals, and needs to be revised for us to achieve bigger gains. Specifically, the bound on perturbation has created a somewhat contrived setting and needs to be relaxed. This has misled us to focus on model classes that are not expressive enough to begin with. Instead, inspired by human vision and the fact that we rely more on robust features such as shape, vertices, and foreground objects than non-robust features such as texture, efforts should be steered towards looking for significantly different classes of models. Maybe instead of narrowing down on imperceptible adversarial perturbations, we should attack a more general problem which is finding architectures that are simultaneously robust to perceptible perturbations, geometric transformations (e.g. rotation, scaling), image distortions (lighting, blur), and more (e.g. occlusion, shadow). Only then we may be able to solve the problem of adversarial vulnerability.

Ali Borji

Test sets are an integral part of evaluating models and gauging progress in object recognition, and more broadly in computer vision and AI. Existing test sets for object recognition, however, suffer from shortcomings such as bias towards the ImageNet characteristics and idiosyncrasies (e.g., ImageNet-V2), being limited to certain types of stimuli (e.g., indoor scenes in ObjectNet), and underestimating the model performance (e.g., ImageNet-A). To mitigate these problems, we introduce a new test set, called D2O, which is sufficiently different from existing test sets. Images are a mix of generated images as well as images crawled from the web. They are diverse, unmodified, and representative of real-world scenarios and cause state-of-the-art models to misclassify them with high confidence. To emphasize generalization, our dataset by design does not come paired with a training set. It contains 8,060 images spread across 36 categories, out of which 29 appear in ImageNet. The best Top-1 accuracy on our dataset is around 60% which is much lower than 91% best Top-1 accuracy on ImageNet. We find that popular vision APIs perform very poorly in detecting objects over D2O categories such as ``faces'', ``cars'', and ``cats''. Our dataset also comes with a ``miscellaneous'' category, over which we test the image tagging models. Overall, our investigations demonstrate that the D2O test set contain a mix of images with varied levels of difficulty and is predictive of the average-case performance of models. It can challenge object recognition models for years to come and can spur more research in this fundamental area.

Ali Borji

Although extensive research has been carried out to evaluate the effectiveness of AI tools and models in detecting deep fakes, the question remains unanswered regarding whether these models can accurately identify genuine images that appear artificial. In this study, as an initial step towards addressing this issue, we have curated a dataset of 510 genuine images that exhibit a fake appearance and conducted an assessment using two AI models. We show that two models exhibited subpar performance when applied to our dataset. Additionally, our dataset can serve as a valuable tool for assessing the ability of deep learning models to comprehend complex visual stimuli. We anticipate that this research will stimulate further discussions and investigations in this area. Our dataset is accessible at https://github.com/aliborji/FLORIDA.

Ali Borji

The study conducted by Shumailov et al. (2024) demonstrates that repeatedly training a generative model on synthetic data leads to model collapse. This finding has generated considerable interest and debate, particularly given that current models have nearly exhausted the available data. In this work, we investigate the effects of fitting a distribution (through Kernel Density Estimation, or KDE) or a model to the data, followed by repeated sampling from it. Our objective is to develop a theoretical understanding of the phenomenon observed by Shumailov et al. (2024). Our results indicate that the outcomes reported are a statistical phenomenon and may be unavoidable.

Ali Borji

Humans rely heavily on shape information to recognize objects. Conversely, convolutional neural networks (CNNs) are biased more towards texture. This is perhaps the main reason why CNNs are vulnerable to adversarial examples. Here, we explore how shape bias can be incorporated into CNNs to improve their robustness. Two algorithms are proposed, based on the observation that edges are invariant to moderate imperceptible perturbations. In the first one, a classifier is adversarially trained on images with the edge map as an additional channel. At inference time, the edge map is recomputed and concatenated to the image. In the second algorithm, a conditional GAN is trained to translate the edge maps, from clean and/or perturbed images, into clean images. Inference is done over the generated image corresponding to the input's edge map. Extensive experiments over 10 datasets demonstrate the effectiveness of the proposed algorithms against FGSM and $\ell_\infty$ PGD-40 attacks. Further, we show that a) edge information can also benefit other adversarial training methods, and b) CNNs trained on edge-augmented inputs are more robust against natural image corruptions such as motion blur, impulse noise and JPEG compression, than CNNs trained solely on RGB images. From a broader perspective, our study suggests that CNNs do not adequately account for image structures that are crucial for robustness. Code is available at:~\url{https://github.com/aliborji/Shapedefense.git}.

Ali Borji

Object detection remains as one of the most notorious open problems in computer vision. Despite large strides in accuracy in recent years, modern object detectors have started to saturate on popular benchmarks raising the question of how far we can reach with deep learning tools and tricks. Here, by employing 2 state-of-the-art object detection benchmarks, and analyzing more than 15 models over 4 large scale datasets, we I) carefully determine the upper bound in AP, which is 91.6% on VOC (test2007), 78.2% on COCO (val2017), and 58.9% on OpenImages V4 (validation), regardless of the IOU threshold. These numbers are much better than the mAP of the best model (47.9% on VOC, and 46.9% on COCO; IOUs=.5:.05:.95), II) characterize the sources of errors in object detectors, in a novel and intuitive way, and find that classification error (confusion with other classes and misses) explains the largest fraction of errors and weighs more than localization and duplicate errors, and III) analyze the invariance properties of models when surrounding context of an object is removed, when an object is placed in an incongruent background, and when images are blurred or flipped vertically. We find that models generate a lot of boxes on empty regions and that context is more important for detecting small objects than larger ones. Our work taps into the tight relationship between object detection and object recognition and offers insights for building better models. Our code is publicly available at https://github.com/aliborji/Deetctionupper bound.git.

Hamed R. Tavakoli, Ali Borji, Esa Rahtu et al.

This paper studies audio-visual deep saliency prediction. It introduces a conceptually simple and effective Deep Audio-Visual Embedding for dynamic saliency prediction dubbed ``DAVE" in conjunction with our efforts towards building an Audio-Visual Eye-tracking corpus named ``AVE". Despite existing a strong relation between auditory and visual cues for guiding gaze during perception, video saliency models only consider visual cues and neglect the auditory information that is ubiquitous in dynamic scenes. Here, we investigate the applicability of audio cues in conjunction with visual ones in predicting saliency maps using deep neural networks. To this end, the proposed model is intentionally designed to be simple. Two baseline models are developed on the same architecture which consists of an encoder-decoder. The encoder projects the input into a feature space followed by a decoder that infers saliency. We conduct an extensive analysis on different modalities and various aspects of multi-model dynamic saliency prediction. Our results suggest that (1) audio is a strong contributing cue for saliency prediction, (2) salient visible sound-source is the natural cause of the superiority of our Audio-Visual model, (3) richer feature representations for the input space leads to more powerful predictions even in absence of more sophisticated saliency decoders, and (4) Audio-Visual model improves over 53.54\% of the frames predicted by the best Visual model (our baseline). Our endeavour demonstrates that audio is an important cue that boosts dynamic video saliency prediction and helps models to approach human performance. The code is available at https://github.com/hrtavakoli/DAVE

Zhaohui Che, Ali Borji, Guangtao Zhai et al.

Data size is the bottleneck for developing deep saliency models, because collecting eye-movement data is very time consuming and expensive. Most of current studies on human attention and saliency modeling have used high quality stereotype stimuli. In real world, however, captured images undergo various types of transformations. Can we use these transformations to augment existing saliency datasets? Here, we first create a novel saliency dataset including fixations of 10 observers over 1900 images degraded by 19 types of transformations. Second, by analyzing eye movements, we find that observers look at different locations over transformed versus original images. Third, we utilize the new data over transformed images, called data augmentation transformation (DAT), to train deep saliency models. We find that label preserving DATs with negligible impact on human gaze boost saliency prediction, whereas some other DATs that severely impact human gaze degrade the performance. These label preserving valid augmentation transformations provide a solution to enlarge existing saliency datasets. Finally, we introduce a novel saliency model based on generative adversarial network (dubbed GazeGAN). A modified UNet is proposed as the generator of the GazeGAN, which combines classic skip connections with a novel center-surround connection (CSC), in order to leverage multi level features. We also propose a histogram loss based on Alternative Chi Square Distance (ACS HistLoss) to refine the saliency map in terms of luminance distribution. Extensive experiments and comparisons over 3 datasets indicate that GazeGAN achieves the best performance in terms of popular saliency evaluation metrics, and is more robust to various perturbations. Our code and data are available at: https://github.com/CZHQuality/Sal-CFS-GAN.

Sen He, Hamed R. Tavakoli, Ali Borji et al.

In this work, we present a novel dataset consisting of eye movements and verbal descriptions recorded synchronously over images. Using this data, we study the differences in human attention during free-viewing and image captioning tasks. We look into the relationship between human attention and language constructs during perception and sentence articulation. We also analyse attention deployment mechanisms in the top-down soft attention approach that is argued to mimic human attention in captioning tasks, and investigate whether visual saliency can help image captioning. Our study reveals that (1) human attention behaviour differs in free-viewing and image description tasks. Humans tend to fixate on a greater variety of regions under the latter task, (2) there is a strong relationship between described objects and attended objects ($97\%$ of the described objects are being attended), (3) a convolutional neural network as feature encoder accounts for human-attended regions during image captioning to a great extent (around $78\%$), (4) soft-attention mechanism differs from human attention, both spatially and temporally, and there is low correlation between caption scores and attention consistency scores. These indicate a large gap between humans and machines in regards to top-down attention, and (5) by integrating the soft attention model with image saliency, we can significantly improve the model's performance on Flickr30k and MSCOCO benchmarks. The dataset can be found at: https://github.com/SenHe/Human-Attention-in-Image-Captioning.

Morteza Moradi, Mohammad Moradi, Simone Palazzo et al.

Salient Object Detection (SOD) remains an essential yet underexplored task in the era of large-scale vision models. Although foundation models like SAM exhibit strong generalization, their potential for SOD is not fully realized, and training or fully fine-tuning them is computationally expensive and prone to overfitting under limited data. To overcome these challenges, we introduce GLASSNet, a Global-Local feature decoding framework that uses SAMv2 as a frozen encoder paired with a lightweight, spatially aware convolutional adapter-reducing learnable encoder parameters by over 97%. To enhance saliency quality, GLASSNet employs a dual-decoder architecture: one decoder captures global, long-range semantics with an expanded receptive field, while the other captures fine local details such as edges and textures. Fusing these complementary cues yields saliency maps that combine global coherence with local precision, producing accurate final masks. Extensive experiments on standard SOD and camouflaged object detection benchmarks show that GLASSNet surpasses state-of-the-art methods, demonstrating the power of frozen foundation models combined with targeted adaptation and global-local decoding.

Morteza Moradi, Mohammad Moradi, Francesco Rundo et al.

Recent advancements in video saliency prediction (VSP) have shown promising performance compared to the human visual system, whose emulation is the primary goal of VSP. However, current state-of-the-art models employ spatio-temporal transformers trained on limited amounts of data, hindering generalizability adaptation to downstream tasks. The benefits of vision foundation models present a potential solution to improve the VSP process. However, adapting image foundation models to the video domain presents significant challenges in modeling scene dynamics and capturing temporal information. To address these challenges, and as the first initiative to design a VSP model based on video foundation models, we introduce SalFoM, a novel encoder-decoder video transformer architecture. Our model employs UnMasked Teacher (UMT) as feature extractor and presents a heterogeneous decoder which features a locality-aware spatio-temporal transformer and integrates local and global spatio-temporal information from various perspectives to produce the final saliency map. Our qualitative and quantitative experiments on the challenging VSP benchmark datasets of DHF1K, Hollywood-2 and UCF-Sports demonstrate the superiority of our proposed model in comparison with the state-of-the-art methods.

Ali Borji

Transformers have emerged as the prevailing standard solution for various AI tasks, including computer vision and natural language processing. The widely adopted Query, Key, and Value formulation (QKV) has played a significant role in this. Nevertheless, no research has examined the essentiality of these three components for transformer performance. Therefore, we conducted an evaluation of the key-value formulation (KV), which generates symmetric attention maps, along with an asymmetric version that incorporates a 2D positional encoding into the attention matrix. Remarkably, this transformer requires fewer parameters and computation than the original one. Through experiments encompassing three task types -- synthetics (such as reversing or sorting a list), vision (mnist or cifar classification), and NLP (character generation and translation) -- we discovered that the KV transformer occasionally outperforms the QKV transformer. However, it also exhibits instances of underperformance compared to QKV, making it challenging to draw a definitive conclusion. Nonetheless, we consider the reported results to be encouraging and anticipate that they may pave the way for more efficient transformers in the future.

Ali Borji

Overparametrization has become a de facto standard in machine learning. Despite numerous efforts, our understanding of how and where overparametrization helps model accuracy and robustness is still limited. To this end, here we conduct an empirical investigation to systemically study and replicate previous findings in this area, in particular the study by Madry et al. Together with this study, our findings support the "universal law of robustness" recently proposed by Bubeck et al. We argue that while critical for robust perception, overparametrization may not be enough to achieve full robustness and smarter architectures e.g. the ones implemented by the human visual cortex) seem inevitable.

Minglang Qiao, Yufan Liu, Mai Xu et al.

Visual and audio events simultaneously occur and both attract attention. However, most existing saliency prediction works ignore the influence of audio and only consider vision modality. In this paper, we propose a multitask learning method for visual-audio saliency prediction and sound source localization on multi-face video by leveraging visual, audio and face information. Specifically, we first introduce a large-scale database of multi-face video in visual-audio condition (MVVA), containing eye-tracking data and sound source annotations. Using this database, we find that sound influences human attention, and conversly attention offers a cue to determine sound source on multi-face video. Guided by these findings, a visual-audio multi-task network (VAM-Net) is introduced to predict saliency and locate sound source. VAM-Net consists of three branches corresponding to visual, audio and face modalities. Visual branch has a two-stream architecture to capture spatial and temporal information. Face and audio branches encode audio signals and faces, respectively. Finally, a spatio-temporal multi-modal graph (STMG) is constructed to model the interaction among multiple faces. With joint optimization of these branches, the intrinsic correlation of the tasks of saliency prediction and sound source localization is utilized and their performance is boosted by each other. Experiments show that the proposed method outperforms 12 state-of-the-art saliency prediction methods, and achieves competitive results in sound source localization.

Yufan Liu, Minglang Qiao, Mai Xu et al.

Recently, video streams have occupied a large proportion of Internet traffic, most of which contain human faces. Hence, it is necessary to predict saliency on multiple-face videos, which can provide attention cues for many content based applications. However, most of multiple-face saliency prediction works only consider visual information and ignore audio, which is not consistent with the naturalistic scenarios. Several behavioral studies have established that sound influences human attention, especially during the speech turn-taking in multiple-face videos. In this paper, we thoroughly investigate such influences by establishing a large-scale eye-tracking database of Multiple-face Video in Visual-Audio condition (MVVA). Inspired by the findings of our investigation, we propose a novel multi-modal video saliency model consisting of three branches: visual, audio and face. The visual branch takes the RGB frames as the input and encodes them into visual feature maps. The audio and face branches encode the audio signal and multiple cropped faces, respectively. A fusion module is introduced to integrate the information from three modalities, and to generate the final saliency map. Experimental results show that the proposed method outperforms 11 state-of-the-art saliency prediction works. It performs closer to human multi-modal attention.

Ali Borji

This work is an update of a previous paper on the same topic published a few years ago. With the dramatic progress in generative modeling, a suite of new quantitative and qualitative techniques to evaluate models has emerged. Although some measures such as Inception Score, Frechet Inception Distance, Precision-Recall, and Perceptual Path Length are relatively more popular, GAN evaluation is not a settled issue and there is still room for improvement. Here, I describe new dimensions that are becoming important in assessing models (e.g. bias and fairness) and discuss the connection between GAN evaluation and deepfakes. These are important areas of concern in the machine learning community today and progress in GAN evaluation can help mitigate them.

Ali Borji

High image resolution is critical to obtain a good performance in many computer vision applications. Computational complexity of CNNs, however, grows significantly with the increase in input image size. Here, we show that it is almost always possible to modify a network such that it achieves higher accuracy at a higher input resolution while having the same number of parameters or/and FLOPS. The idea is similar to the EfficientNet paper but instead of optimizing network width, depth and resolution simultaneously, here we focus only on input resolution. This makes the search space much smaller which is more suitable for low computational budget regimes. More importantly, by controlling for the number of model parameters (and hence model capacity), we show that the additional benefit in accuracy is indeed due to the higher input resolution. Preliminary empirical investigation over MNIST, Fashion MNIST, and CIFAR10 datasets demonstrates the efficiency of the proposed approach.

Ali Borji

Deep object recognition models have been very successful over benchmark datasets such as ImageNet. How accurate and robust are they to distribution shifts arising from natural and synthetic variations in datasets? Prior research on this problem has primarily focused on ImageNet variations (e.g., ImageNetV2, ImageNet-A). To avoid potential inherited biases in these studies, we take a different approach. Specifically, we reanalyze the ObjectNet dataset recently proposed by Barbu et al. containing objects in daily life situations. They showed a dramatic performance drop of the state of the art object recognition models on this dataset. Due to the importance and implications of their results regarding the generalization ability of deep models, we take a second look at their analysis. We find that applying deep models to the isolated objects, rather than the entire scene as is done in the original paper, results in around 20-30% performance improvement. Relative to the numbers reported in Barbu et al., around 10-15% of the performance loss is recovered, without any test time data augmentation. Despite this gain, however, we conclude that deep models still suffer drastically on the ObjectNet dataset. We also investigate the robustness of models against synthetic image perturbations such as geometric transformations (e.g., scale, rotation, translation), natural image distortions (e.g., impulse noise, blur) as well as adversarial attacks (e.g., FGSM and PGD-5). Our results indicate that limiting the object area as much as possible (i.e., from the entire image to the bounding box to the segmentation mask) leads to consistent improvement in accuracy and robustness.

Saman Zabihi, Hamed Rezazadegan Tavakoli, Ali Borji

Saliency computation models aim to imitate the attention mechanism in the human visual system. The application of deep neural networks for saliency prediction has led to a drastic improvement over the last few years. However, deep models have a high number of parameters which makes them less suitable for real-time applications. Here we propose a compact yet fast model for real-time saliency prediction. Our proposed model consists of a modified U-net architecture, a novel fully connected layer, and central difference convolutional layers. The modified U-Net architecture promotes compactness and efficiency. The novel fully-connected layer facilitates the implicit capturing of the location-dependent information. Using the central difference convolutional layers at different scales enables capturing more robust and biologically motivated features. We compare our model with state of the art saliency models using traditional saliency scores as well as our newly devised scheme. Experimental results over four challenging saliency benchmark datasets demonstrate the effectiveness of our approach in striking a balance between accuracy and speed. Our model can be run in real-time which makes it appealing for edge devices and video processing.

Yingjie Zhai, Deng-Ping Fan, Jufeng Yang et al.

Multi-level feature fusion is a fundamental topic in computer vision. It has been exploited to detect, segment and classify objects at various scales. When multi-level features meet multi-modal cues, the optimal feature aggregation and multi-modal learning strategy become a hot potato. In this paper, we leverage the inherent multi-modal and multi-level nature of RGB-D salient object detection to devise a novel cascaded refinement network. In particular, first, we propose to regroup the multi-level features into teacher and student features using a bifurcated backbone strategy (BBS). Second, we introduce a depth-enhanced module (DEM) to excavate informative depth cues from the channel and spatial views. Then, RGB and depth modalities are fused in a complementary way. Our architecture, named Bifurcated Backbone Strategy Network (BBS-Net), is simple, efficient, and backbone-independent. Extensive experiments show that BBS-Net significantly outperforms eighteen SOTA models on eight challenging datasets under five evaluation measures, demonstrating the superiority of our approach ($\sim 4 \%$ improvement in S-measure $vs.$ the top-ranked model: DMRA-iccv2019). In addition, we provide a comprehensive analysis on the generalization ability of different RGB-D datasets and provide a powerful training set for future research.

Ali Borji

Deep learning has come a long way and has enjoyed an unprecedented success. Despite high accuracy, however, deep models are brittle and are easily fooled by imperceptible adversarial perturbations. In contrast to common inference-time attacks, Backdoor (\aka Trojan) attacks target the training phase of model construction, and are extremely difficult to combat since a) the model behaves normally on a pristine testing set and b) the augmented perturbations can be minute and may only affect few training samples. Here, I propose a new method to tell whether a model has been subject to a backdoor attack. The idea is to generate adversarial examples, targeted or untargeted, using conventional attacks such as FGSM and then feed them back to the classifier. By computing the statistics (here simply mean maps) of the images in different categories and comparing them with the statistics of a reference model, it is possible to visually locate the perturbed regions and unveil the attack.

Ali Borji

I introduce a very simple method to defend against adversarial examples. The basic idea is to raise the slope of the ReLU function at the test time. Experiments over MNIST and CIFAR-10 datasets demonstrate the effectiveness of the proposed defense against a number of strong attacks in both untargeted and targeted settings. While perhaps not as effective as the state of the art adversarial defenses, this approach can provide insights to understand and mitigate adversarial attacks. It can also be used in conjunction with other defenses.

Ali Borji

Recently, Barbu et al introduced a dataset called ObjectNet which includes objects in daily life situations. They showed a dramatic performance drop of the state of the art object recognition models on this dataset. Due to the importance and implications of their results regarding generalization ability of deep models, we take a second look at their findings. We highlight a major problem with their work which is applying object recognizers to the scenes containing multiple objects rather than isolated objects. The latter results in around 20-30% performance gain using our code. Compared with the results reported in the ObjectNet paper, we observe that around 10-15 % of the performance loss can be recovered, without any test time data augmentation. In accordance with Barbu et al.'s conclusions, however, we also conclude that deep models suffer drastically on this dataset. Thus, we believe that ObjectNet remains a challenging dataset for testing the generalization power of models beyond datasets on which they have been trained.

Ali Borji, Sikun Lin

A white noise analysis of modern deep neural networks is presented to unveil their biases at the whole network level or the single neuron level. Our analysis is based on two popular and related methods in psychophysics and neurophysiology namely classification images and spike triggered analysis. These methods have been widely used to understand the underlying mechanisms of sensory systems in humans and monkeys. We leverage them to investigate the inherent biases of deep neural networks and to obtain a first-order approximation of their functionality. We emphasize on CNNs since they are currently the state of the art methods in computer vision and are a decent model of human visual processing. In addition, we study multi-layer perceptrons, logistic regression, and recurrent neural networks. Experiments over four classic datasets, MNIST, Fashion-MNIST, CIFAR-10, and ImageNet, show that the computed bias maps resemble the target classes and when used for classification lead to an over twofold performance than the chance level. Further, we show that classification images can be used to attack a black-box classifier and to detect adversarial patch attacks. Finally, we utilize spike triggered averaging to derive the filters of CNNs and explore how the behavior of a network changes when neurons in different layers are modulated. Our effort illustrates a successful example of borrowing from neurosciences to study ANNs and highlights the importance of cross-fertilization and synergy across machine learning, deep learning, and computational neuroscience.

Ali Borji, Seyed Mehdi Iranmanesh

Object detection remains as one of the most notorious open problems in computer vision. Despite large strides in accuracy in recent years, modern object detectors have started to saturate on popular benchmarks raising the question of how far we can reach with deep learning tools and tricks. Here, by employing 2 state-of-the-art object detection benchmarks, and analyzing more than 15 models over 4 large scale datasets, we I) carefully determine the upperbound in AP, which is 91.6% on VOC (test2007), 78.2% on COCO (val2017), and 58.9% on OpenImages V4 (validation), regardless of the IOU. These numbers are much better than the mAP of the best model1 (47.9% on VOC, and 46.9% on COCO; IOUs=.5:.95), II) characterize the sources of errors in object detectors, in a novel and intuitive way, and find that classification error (confusion with other classes and misses) explains the largest fraction of errors and weighs more than localization and duplicate errors, and III) analyze the invariance properties of models when surrounding context of an object is removed, when an object is placed in an incongruent background, and when images are blurred or flipped vertically. We find that models generate boxes on empty regions and that context is more important for detecting small objects than larger ones. Our work taps into the tight relationship between recognition and detection and offers insights for building better models.

Zhaohui Che, Ali Borji, Guangtao Zhai et al.

Deep neural networks are vulnerable to adversarial attacks.

Hamed R. Tavakoli, Esa Rahtu, Juho Kannala et al.

This paper digs deeper into factors that influence egocentric gaze. Instead of training deep models for this purpose in a blind manner, we propose to inspect factors that contribute to gaze guidance during daily tasks. Bottom-up saliency and optical flow are assessed versus strong spatial prior baselines. Task-specific cues such as vanishing point, manipulation point, and hand regions are analyzed as representatives of top-down information. We also look into the contribution of these factors by investigating a simple recurrent neural model for ego-centric gaze prediction. First, deep features are extracted for all input video frames. Then, a gated recurrent unit is employed to integrate information over time and to predict the next fixation. We also propose an integrated model that combines the recurrent model with several top-down and bottom-up cues. Extensive experiments over multiple datasets reveal that (1) spatial biases are strong in egocentric videos, (2) bottom-up saliency models perform poorly in predicting gaze and underperform spatial biases, (3) deep features perform better compared to traditional features, (4) as opposed to hand regions, the manipulation point is a strong influential cue for gaze prediction, (5) combining the proposed recurrent model with bottom-up cues, vanishing points and, in particular, manipulation point results in the best gaze prediction accuracy over egocentric videos, (6) the knowledge transfer works best for cases where the tasks or sequences are similar, and (7) task and activity recognition can benefit from gaze prediction. Our findings suggest that (1) there should be more emphasis on hand-object interaction and (2) the egocentric vision community should consider larger datasets including diverse stimuli and more subjects.

Zhaohui Che, Ali Borji, Guangtao Zhai et al.

Currently, a plethora of saliency models based on deep neural networks have led great breakthroughs in many complex high-level vision tasks (e.g. scene description, object detection). The robustness of these models, however, has not yet been studied. In this paper, we propose a sparse feature-space adversarial attack method against deep saliency models for the first time. The proposed attack only requires a part of the model information, and is able to generate a sparser and more insidious adversarial perturbation, compared to traditional image-space attacks. These adversarial perturbations are so subtle that a human observer cannot notice their presences, but the model outputs will be revolutionized. This phenomenon raises security threats to deep saliency models in practical applications. We also explore some intriguing properties of the feature-space attack, e.g. 1) the hidden layers with bigger receptive fields generate sparser perturbations, 2) the deeper hidden layers achieve higher attack success rates, and 3) different loss functions and different attacked layers will result in diverse perturbations. Experiments indicate that the proposed method is able to successfully attack different model architectures across various image scenes.

Sen He, Hamed R. Tavakoli, Ali Borji et al.

Recently, data-driven deep saliency models have achieved high performance and have outperformed classical saliency models, as demonstrated by results on datasets such as the MIT300 and SALICON. Yet, there remains a large gap between the performance of these models and the inter-human baseline. Some outstanding questions include what have these models learned, how and where they fail, and how they can be improved. This article attempts to answer these questions by analyzing the representations learned by individual neurons located at the intermediate layers of deep saliency models. To this end, we follow the steps of existing deep saliency models, that is borrowing a pre-trained model of object recognition to encode the visual features and learning a decoder to infer the saliency. We consider two cases when the encoder is used as a fixed feature extractor and when it is fine-tuned, and compare the inner representations of the network. To study how the learned representations depend on the task, we fine-tune the same network using the same image set but for two different tasks: saliency prediction versus scene classification. Our analyses reveal that: 1) some visual regions (e.g. head, text, symbol, vehicle) are already encoded within various layers of the network pre-trained for object recognition, 2) using modern datasets, we find that fine-tuning pre-trained models for saliency prediction makes them favor some categories (e.g. head) over some others (e.g. text), 3) although deep models of saliency outperform classical models on natural images, the converse is true for synthetic stimuli (e.g. pop-out search arrays), an evidence of significant difference between human and data-driven saliency models, and 4) we confirm that, after-fine tuning, the change in inner-representations is mostly due to the task and not the domain shift in the data.

Mohamed Elfeki, Ali Borji

To automatically produce a brief yet expressive summary of a long video, an automatic algorithm should start by resembling the human process of summary generation. Prior work proposed supervised and unsupervised algorithms to train models for learning the underlying behavior of humans by increasing modeling complexity or craft-designing better heuristics to simulate human summary generation process. In this work, we take a different approach by analyzing a major cue that humans exploit for the summary generation; the nature and intensity of actions. We empirically observed that a frame is more likely to be included in human-generated summaries if it contains a substantial amount of deliberate motion performed by an agent, which is referred to as actionness. Therefore, we hypothesize that learning to automatically generate summaries involves an implicit knowledge of actionness estimation and ranking. We validate our hypothesis by running a user study that explores the correlation between human-generated summaries and actionness ranks. We also run a consensus and behavioral analysis between human subjects to ensure reliable and consistent results. The analysis exhibits a considerable degree of agreement among subjects within obtained data and verifying our initial hypothesis. Based on the study findings, we develop a method to incorporate actionness data to explicitly regulate a learning algorithm that is trained for summary generation. We assess the performance of our approach to four summarization benchmark datasets and demonstrate an evident advantage compared to state-of-the-art summarization methods.

Mohamed Elfeki, Liqiang Wang, Ali Borji

With vast amounts of video content being uploaded to the Internet every minute, video summarization becomes critical for efficient browsing, searching, and indexing of visual content. Nonetheless, the spread of social and egocentric cameras creates an abundance of sparse scenarios captured by several devices, and ultimately required to be jointly summarized. In this paper, we discuss the problem of summarizing videos recorded independently by several dynamic cameras that intermittently share the field of view. We present a robust framework that (a) identifies a diverse set of important events among moving cameras that often are not capturing the same scene, and (b) selects the most representative view(s) at each event to be included in a universal summary. Due to the lack of an applicable alternative, we collected a new multi-view egocentric dataset, Multi-Ego. Our dataset is recorded simultaneously by three cameras, covering a wide variety of real-life scenarios. The footage is annotated by multiple individuals under various summarization configurations, with a consensus analysis ensuring a reliable ground truth. We conduct extensive experiments on the compiled dataset in addition to three other standard benchmarks that show the robustness and the advantage of our approach in both supervised and unsupervised settings. Additionally, we show that our approach learns collectively from data of varied number-of-views and orthogonal to other summarization methods, deeming it scalable and generic.

Mohamed Elfeki, Krishna Regmi, Shervin Ardeshir et al.

First-person (egocentric) and third person (exocentric) videos are drastically different in nature. The relationship between these two views have been studied in recent years, however, it has yet to be fully explored. In this work, we introduce two datasets (synthetic and natural/real) containing simultaneously recorded egocentric and exocentric videos. We also explore relating the two domains (egocentric and exocentric) in two aspects. First, we synthesize images in the egocentric domain from the exocentric domain using a conditional generative adversarial network (cGAN). We show that with enough training data, our network is capable of hallucinating how the world would look like from an egocentric perspective, given an exocentric video. Second, we address the cross-view retrieval problem across the two views. Given an egocentric query frame (or its momentary optical flow), we retrieve its corresponding exocentric frame (or optical flow) from a gallery set. We show that using synthetic data could be beneficial in retrieving real data. We show that performing domain adaptation from the synthetic domain to the natural/real domain, is helpful in tasks such as retrieval. We believe that the presented datasets and the proposed baselines offer new opportunities for further research in this direction. The code and dataset are publicly available.

Ali Borji, Hamed R. Tavakoli, Zoya Bylinskii

In this review, we examine the recent progress in saliency prediction and proposed several avenues for future research. In spite of tremendous efforts and huge progress, there is still room for improvement in terms finer-grained analysis of deep saliency models, evaluation measures, datasets, annotation methods, cognitive studies, and new applications. This chapter will appear in Encyclopedia of Computational Neuroscience.

Zhaohui Che, Ali Borji, Guangtao Zhai et al.

Most of current studies on human gaze and saliency modeling have used high-quality stimuli. In real world, however, captured images undergo various types of distortions during the whole acquisition, transmission, and displaying chain. Some distortion types include motion blur, lighting variations and rotation. Despite few efforts, influences of ubiquitous distortions on visual attention and saliency models have not been systematically investigated. In this paper, we first create a large-scale database including eye movements of 10 observers over 1900 images degraded by 19 types of distortions. Second, by analyzing eye movements and saliency models, we find that: a) observers look at different locations over distorted versus original images, and b) performances of saliency models are drastically hindered over distorted images, with the maximum performance drop belonging to Rotation and Shearing distortions. Finally, we investigate the effectiveness of different distortions when serving as data augmentation transformations. Experimental results verify that some useful data augmentation transformations which preserve human gaze of reference images can improve deep saliency models against distortions, while some invalid transformations which severely change human gaze will degrade the performance.

Ali Borji

Visual saliency models have enjoyed a big leap in performance in recent years, thanks to advances in deep learning and large scale annotated data. Despite enormous effort and huge breakthroughs, however, models still fall short in reaching human-level accuracy. In this work, I explore the landscape of the field emphasizing on new deep saliency models, benchmarks, and datasets. A large number of image and video saliency models are reviewed and compared over two image benchmarks and two large scale video datasets. Further, I identify factors that contribute to the gap between models and humans and discuss remaining issues that need to be addressed to build the next generation of more powerful saliency models. Some specific questions that are addressed include: in what ways current models fail, how to remedy them, what can be learned from cognitive studies of attention, how explicit saliency judgments relate to fixations, how to conduct fair model comparison, and what are the emerging applications of saliency models.

Krishna Regmi, Ali Borji

We address the problem of generating images across two drastically different views, namely ground (street) and aerial (overhead) views. Image synthesis by itself is a very challenging computer vision task and is even more so when generation is conditioned on an image in another view. Due the difference in viewpoints, there is small overlapping field of view and little common content between these two views. Here, we try to preserve the pixel information between the views so that the generated image is a realistic representation of cross view input image. For this, we propose to use homography as a guide to map the images between the views based on the common field of view to preserve the details in the input image. We then use generative adversarial networks to inpaint the missing regions in the transformed image and add realism to it. Our exhaustive evaluation and model comparison demonstrate that utilizing geometry constraints adds fine details to the generated images and can be a better approach for cross view image synthesis than purely pixel based synthesis methods.

Aidean Sharghi, Ali Borji, Chengtao Li et al.

It is now much easier than ever before to produce videos. While the ubiquitous video data is a great source for information discovery and extraction, the computational challenges are unparalleled. Automatically summarizing the videos has become a substantial need for browsing, searching, and indexing visual content. This paper is in the vein of supervised video summarization using sequential determinantal point process (SeqDPP), which models diversity by a probabilistic distribution. We improve this model in two folds. In terms of learning, we propose a large-margin algorithm to address the exposure bias problem in SeqDPP. In terms of modeling, we design a new probabilistic distribution such that, when it is integrated into SeqDPP, the resulting model accepts user input about the expected length of the summary. Moreover, we also significantly extend a popular video summarization dataset by 1) more egocentric videos, 2) dense user annotations, and 3) a refined evaluation scheme. We conduct extensive experiments on this dataset (about 60 hours of videos in total) and compare our approach to several competitive baselines.