Class-agnostic counting (CAC) aims to estimate the number of objects in images without being restricted to predefined categories. However, while current exemplar-based CAC methods offer flexibility at inference time, they still heavily rely on labeled data for training, which limits scalability and generalization to many downstream use cases. In this paper, we introduce CountingDINO, the first training-free exemplar-based CAC framework that exploits a fully unsupervised feature extractor. Specifically, our approach employs self-supervised vision-only backbones to extract object-aware features, and it eliminates the need for annotated data throughout the entire proposed pipeline. At inference time, we extract latent object prototypes via ROI-Align from DINO features and use them as convolutional kernels to generate similarity maps. These are then transformed into density maps through a simple yet effective normalization scheme. We evaluate our approach on the FSC-147 and CARPK benchmarks, where we consistently outperform a baseline based on an SOTA unsupervised object detector under the same label- and training-free setting. Additionally, we achieve competitive results – and in some cases surpass – training-free methods that rely on supervised backbones, non-training-free unsupervised methods, as well as several fully supervised SOTA approaches. This demonstrates that label- and training-free CAC can be both scalable and effective. Website: https://lorebianchi98.github.io/CountingDINO/.

Visual object counting has recently shifted towards class-agnostic counting (CAC), which addresses the challenge of counting objects across arbitrary categories—a crucial capability for flexible and generalizable counting systems. Unlike humans, who effortlessly identify and count objects from diverse categories without prior knowledge, most existing counting methods are restricted to enumerating instances of known classes, requiring extensive labeled datasets for training and struggling in open-vocabulary settings. In contrast, CAC aims to count objects belonging to classes never seen during training, operating in a few-shot setting. In this paper, we present the first comprehensive review of CAC methodologies. We propose a taxonomy to categorize CAC approaches into three paradigms based on how target object classes can be specified: reference-based, reference-less, and open-world text-guided. Reference-based approaches achieve state-of-the-art performance by relying on exemplar-guided mechanisms. Reference-less methods eliminate exemplar dependency by leveraging inherent image patterns. Finally, open-world text-guided methods use vision-language models, enabling object class descriptions via textual prompts, offering a flexible and promising solution. Based on this taxonomy, we provide an overview of 30 CAC architectures and report their performance on gold-standard benchmarks, discussing key strengths and limitations. Specifically, we present results on the FSC-147 dataset, setting a leaderboard using gold-standard metrics, and on the CARPK dataset to assess generalization capabilities. Finally, we offer a critical discussion of persistent challenges, such as annotation dependency and generalization, alongside future directions. We believe this survey offers a valuable resource for researchers, capturing the evolution of CAC and providing insights to guide future developments in the field. • Provides the first comprehensive survey on class-agnostic object counting (CAC). • Proposes a taxonomy: reference-based, reference-less, and open-world text-guided. • Benchmarks 30 CAC methods on FSC-147 and CARPK, setting a leaderboard. • Discusses trade-offs between accuracy, flexibility, and annotation requirements. • Identifies open challenges and future directions for generalizable counting.

Deepfake detectors are typically trained on large sets of pristine and generated images, resulting in limited generalization capacity; they excel at identifying deepfakes created through methods encountered during training but struggle with those generated by unknown techniques.This paper introduces a learning approach aimed at significantly enhancing the generalization capabilities of deepfake detectors. Our method takes inspiration from the unique "fingerprints" that image generation processes consistently introduce into the frequency domain. These fingerprints manifest as structured and distinctly recognizable frequency patterns. We propose to train detectors using only pristine images injecting in part of them crafted frequency patterns, simulating the effects of various deepfake generation techniques without being specific to any. These synthetic patterns are based on generic shapes, grids, or auras.We evaluated our approach using diverse architectures across 25 different generation methods. The models trained with our approach were able to perform state-of-the-art deepfake detection, demonstrating also superior generalization capabilities in comparison with previous methods. Indeed, they are untied to any specific generation technique and can effectively identify deepfakes regardless of how they were made.The code to use the proposed approach and reproduce the presented experiments is available at: https://github.com/davide-coccomini/Deepfake-Detection-without-Deepfakes-Generalization-via-Synthetic-Frequency-Patterns-Injection

We propose a novel bio-inspired two-stage semi-supervised learning approach for training semantic segmentation models based on downsampling-upsampling architectures. The first stage does not use backpropagation. Rather, it exploits the Hebbian principle “fire together, wire together” as a local learning rule for updating the weights of both convolutional and transpose-convolutional layers, allowing unsupervised discovery of data features. In the second stage, the model is fine-tuned with standard backpropagation on a small subset of labeled data. We evaluate our methodology through experiments conducted on several widely used biomedical datasets, deeming that this domain is paramount in computer vision and is notably impacted by data scarcity. Results show that our proposed method outperforms SOTA approaches across different levels of label availability. Furthermore, we show that using our unsupervised stage to initialize the SOTA approaches leads to performance improvements. The code to replicate our experiments can be found at https://tinyurl.com/hebbian-semantic-segmentation .

Open-vocabulary 3D scene understanding, i.e., recognizing and classifying objects in 3D scenes without being limited to a predefined set of classes, is a foundational task for robotics and extended reality applications. Current leading methods often rely on 2D foundation models to extract semantics, then projected in 3D. This paper investigates the viability of a purely 3D-native pipeline, thereby eliminating dependencies on 2D models and reprojections. We systematically explored various architectural combinations using established 3D components. However, our extensive experiments on benchmark datasets reveal significant performance limitations with this direct 3D-native approach, with performance metrics falling short of expectations. Rather than a simple failure, these outcomes provide critical insights into the current deficiencies of existing 3D models when cascaded for complex open-vocabulary tasks. We highlight the lessons learned, identify the pipeline's limitations (e.g., segmenter-encoder domain gap, robustness to imperfect segmentations), and posit future research directions. We argue that a fundamental rethinking of model design and interplay is necessary to realize the potential of truly 3D-native open-vocabulary understanding.

Open-Vocabulary Segmentation (OVS) aims at segmenting images from free-form textual concepts without predefined training classes. While existing vision-language models such as CLIP can generate segmentation masks by leveraging coarse spatial information from Vision Transformers, they face challenges in spatial localization due to their global alignment of image and text features. Conversely, self-supervised visual models like DINO excel in fine-grained visual encoding but lack integration with language. To bridge this gap, we present Talk2DINO, a novel hybrid approach that combines the spatial accuracy of DINOv2 with the language understanding of CLIP. Our approach aligns the textual embeddings of CLIP to the patch-level features of DINOv2 through a learned mapping function without the need to fine-tune the underlying backbones. At training time, we exploit the attention maps of DINOv2 to selectively align local visual patches with textual embeddings. We show that the powerful semantic and localization abilities of Talk2DINO can enhance the segmentation process, resulting in more natural and less noisy segmentations, and that our approach can also effectively distinguish foreground objects from the background. Experimental results demonstrate that Talk2DINO achieves state-of-the-art performance across several unsupervised OVS benchmarks. Source code and models are publicly available at: https://lorebianchi98.github.io/Talk2DINO/.

Recently, object counting has shifted towards classagnostic counting (CAC), which counts instances of arbitrary object classes never seen during model training. With advancements in robust vision-and-language foundation models, there is a growing interest in prompt-based CAC, where object categories are specified using natural language. However, we identify significant limitations in current benchmarks for evaluating this task, which hinder both accurate assessment and the development of more effective solutions. Specifically, we argue that the current evaluation protocols do not measure the ability of the model to understand which object has to be counted. This is due to two main factors: (i) the shortcomings of CAC datasets, which primarily consist of images containing objects from a single class, and (ii) the limitations of current counting performance evaluators, which are based on traditional class-specific counting and focus solely on counting errors. To fill this gap, we introduce the Prompt-Aware Counting (PrACo) benchmark. It comprises two targeted tests coupled with evaluation metrics specifically designed to quantitatively measure the robustness and trustworthiness of existing prompt-based CAC models. We evaluate state-of-the-art methods and demonstrate that, although some achieve impressive results on standard class-specific counting metrics, they exhibit a significant deficiency in understanding the input prompt, indicating the need for more careful training procedures or revised designs. The code for reproducing our results is available at https://github.com/ciampluca/PrACo.

Animal behavior serves as a reliable indicator of the adaptation of organisms to their environment and their overall well-being. Through rigorous observation of animal actions and interactions, researchers and observers can glean valuable insights into diverse facets of their lives, encompassing health, social dynamics, ecological relationships, and neuroethological dimensions. Although state-of-the-art deep learning models have demonstrated remarkable accuracy in classifying various forms of animal data, their adoption in animal behavior studies remains limited. This survey article endeavors to comprehensively explore deep learning architectures and strategies applied to the identification of animal behavior, spanning auditory, visual, and audiovisual methodologies. The survey categorizes techniques into pose estimation-based and non-pose estimation-based methods, analyzing their applications, effectiveness, and limitations. Furthermore, the manuscript scrutinizes extant animal behavior datasets, offering a detailed examination of the principal challenges confronting this research domain. The article culminates in a comprehensive discussion of key research directions within deep learning that hold potential for advancing the field of animal behavior studies.

In this article, we introduce the Game Boy Learning Environment (GLE), an innovative suite based on Nintendo Game Boy games, crafted to advance and evaluate deep reinforcement learning algorithms on rich and varied gameplay tasks. GLE offers a comprehensive selection of eleven Game Boy environments, spanning nine distinct titles. These environments represent a significant leap in complexity compared to previous endeavors, like the Arcade Learning Environment, presenting challenges for reinforcement learning like intricate long-term planning, strategic foresight, and hierarchical decision-making, posing substantial difficulties even for proficient human players. We delineate the spectrum of available environments and furnish initial baseline results obtained through the development and assessment of intelligent agents, employing established AI methodologies to address individual levels or subtasks within these environments. All software is openly accessible to the public viahttps://github.com/edofazza/GameBoyLearningEnvironment.

Pose-estimation methods enable extracting human motion from common videos in the structured form of 3D skeleton sequences. Despite great application opportunities, effective content-based access to such spatio-temporal motion data is a challenging problem. In this article, we focus on the recently introduced text-motion retrieval tasks, which aim to search for database motions that are the most relevant to a specified natural language textual description ( text-to-motion ) and vice-versa ( motion-to-text ). Despite recent efforts to explore these promising avenues, a primary challenge remains the insufficient data available to train robust text-motion models effectively. To address this issue, we propose to investigate joint-dataset learning—where we train on multiple text-motion datasets simultaneously—together with the introduction of a Cross-Consistent Contrastive Loss (CCCL) function, which regularizes the learned text-motion common space by imposing uni-modal constraints that augment the representation ability of the trained network. To learn a proper motion representation, we also introduce a transformer-based motion encoder, called MoT++, which employs spatio-temporal attention to process skeleton data sequences. We demonstrate the benefits of the proposed approaches on the widely used KIT Motion Language and HumanML3D datasets, including also some results on the recent Motion-X dataset. We perform detailed experimentation on joint-dataset learning and cross-dataset scenarios, showing the effectiveness of each introduced module in a carefully conducted ablation study and, in turn, pointing out the limitations of state-of-the-art methods. The code for reproducing our results is available here: https://github.com/mesnico/MOTpp .

VISIONE is a video search system that integrates multiple search functionalities, allowing users to search for video segments using textual and visual queries, complemented by temporal search capabilities. It exploits state-of-the-art Artificial Intelligence approaches for visual content analysis and highly efficient indexing techniques to ensure fast response and scalability. In the recently concluded Video Browser Showdown (VBS2024) - a well-established international competition in interactive video retrieval - VISIONE ranked first and scored as the best interactive video search system in four out of seven tasks carried out in the competition.This paper provides an overview of the VISIONE system, emphasizing the improvements made to the system in the last year to improve its usability for novice users. A demonstration video showcasing the system's capabilities across 2,300 hours of diverse video content is available online, as well as a simplified demo of VISIONE.

Modern applications increasingly demand flexible computer vision models that adapt to novel concepts not encountered during training. This necessity is pivotal in emerging domains like extended reality, robotics, and autonomous driving, which require the ability to respond to open-world stimuli. A key ingredient is the ability to identify objects based on free-form textual queries defined at inference time - a task known as open-vocabulary object detection. Multimodal backbones like CLIP are the main enabling technology for current open-world perception solutions. Despite performing well on generic queries, recent studies highlighted limitations on the fine-grained recognition capabilities in open-vocabulary settings - i.e., for distinguishing subtle object features like color, shape, and material. In this paper, we perform a detailed examination of these openvocabulary object recognition limitations to find the root cause. We evaluate the performance of CLIP, the most commonly used vision-language backbone, against a fine-grained objectmatching benchmark, revealing interesting analogies between the limitations of open-vocabulary object detectors and their backbones. Experiments suggest that the lack of fine-grained understanding is caused by the poor separability of object characteristics in the CLIP latent space. Therefore, we try to understand whether fine-grained knowledge is present in CLIP embeddings but not exploited at inference time due, for example, to the unsuitability of the cosine similarity matching function, which may discard important object characteristics. Our preliminary experiments show that simple CLIP latent-space re-projections help separate fine-grained concepts, paving the way towards the development of backbones inherently able to process fine-grained details. The code for reproducing these experiments is available at https://github.com/lorebianchi98/FG-CLIP.

Recent advancements in large vision-language models enabled visual object detection in open-vocabulary scenar-ios, where object classes are defined in free-text formats during inference. In this paper, we aim to probe the state-of-the-art methods for open-vocabulary object detection to determine to what extent they understand finegrained prop-erties of objects and their parts. To this end, we intro-duce an evaluation protocol based on dynamic vocabulary generation to test whether models detect, discern, and as-sign the correct fine-grained description to objects in the presence of hard-negative classes. We contribute with a benchmark suite of increasing difficulty and probing dif-ferent properties like color, pattern, and material. We fur-ther enhance our investigation by evaluating several state-of-the-art open-vocabulary object detectors using the proposed protocol and find that most existing solutions, which shine in standard open-vocabulary benchmarks, struggle to accurately capture and distinguish finer object details. We conclude the paper by highlighting the limitations of current methodologies and exploring promising research directions to overcome the discovered drawbacks. Data and code are available at https://lorebianchi98.github.io/FG-OVD/.

Recent work on sample efficient training of Deep Neural Networks (DNNs) proposed a semi-supervised methodology based on biologically inspired Hebbian learning, combined with traditional backprop-based training. Promising results were achieved on various computer vision benchmarks, in scenarios of scarce labeled data availability. However, current Hebbian learning solutions can hardly address large-scale scenarios due to their demanding computational cost. In order to tackle this limitation, in this contribution, we investigate a novel solution, named FastHebb (FH), based on the reformulation of Hebbian learning rules in terms of matrix multiplications, which can be executed more efficiently on GPU. Starting from Soft-Winner-Takes-All (SWTA) and Hebbian Principal Component Analysis (HPCA) learning rules, we formulate their improved FH versions: SWTA-FH and HPCA-FH. We experimentally show that the proposed approach accelerates training speed up to 70 times, allowing us to gracefully scale Hebbian learning experiments on large datasets and network architectures such as ImageNet and VGG.

Abstract With the increasing importance of multimedia and multilingual data in online encyclopedias, novel methods are needed to fill domain gaps and automatically connect different modalities for increased accessibility. For example, Wikipedia is composed of millions of pages written in multiple languages. Images, when present, often lack textual context, thus remaining conceptually floating and harder to find and manage. In this work, we tackle the novel task of associating images from Wikipedia pages with the correct caption among a large pool of available ones written in multiple languages, as required by the image-caption matching Kaggle challenge organized by the Wikimedia Foundation. A system able to perform this task would improve the accessibility and completeness of the underlying multi-modal knowledge graph in online encyclopedias. We propose a cascade of two models powered by the recent Transformer networks able to efficiently and effectively infer a relevance score between the query image data and the captions. We verify through extensive experiments that the proposed cascaded approach effectively handles a large pool of images and captions while maintaining bounded the overall computational complexity at inference time. With respect to other approaches in the challenge leaderboard, we can achieve remarkable improvements over the previous proposals (+8% in nDCG@5 with respect to the sixth position) with constrained resources. The code is publicly available at https://tinyurl.com/wiki-imcap .

In this paper, we present MINTIME, a video deepfake detection method that effectively captures spatial and temporal inconsistencies in videos that depict multiple individuals and varying face sizes. Unlike previous approaches that either employ simplistic a-posteriori aggregation schemes, i.e., averaging or max operations, or only focus on the largest face in the video, our proposed method learns to accurately detect spatio-temporal inconsistencies across multiple identities in a video through a Spatio-Temporal Transformer combined with a Convolutional Neural Network backbone. This is achieved through an Identity-aware Attention mechanism that applies a masking operation on the face sequence to process each identity independently, which enables effective video-level aggregation. Furthermore, our system incorporates two novel embedding schemes: (i) the Temporal Coherent Positional Embedding, which encodes the temporal information of the face sequences of each identity, and (ii) the Size Embedding, which captures the relative sizes of the faces to the video frames. MINTIME achieves state-of-the-art performance on the ForgeryNet dataset, with a remarkable improvement of up to 14% AUC in videos containing multiple people. Moreover, it demonstrates very robust generalization capabilities in cross-forgery and cross-dataset settings. The code is publicly available at: https://github.com/davide-coccomini/MINTIME-Multi-Ident ity-size-iNvariant-TIMEsformer-for-Video-Deepfake-Detection.

This paper presents a revised version of the VISIONE video retrieval system, which offers a wide range of search functionalities, including free text search, spatial color and object search, visual and semantic similarity search, and temporal search. The system is designed to ensure scalability using advanced indexing techniques and effectiveness using cutting-edge Artificial Intelligence technology for visual content analysis. VISIONE was the runner-up in the 2023 Video Browser Showdown competition, demonstrating its comprehensive video retrieval capabilities. In this paper, we detail the improvements made to the search and browsing interface to enhance its usability for non-expert users. A demonstration video of our system with the restyled interface, showcasing its capabilities on over 2,300 hours of diverse video content, is available online at https://youtu.be/srD3TCUkMSg.

Trends and opinion mining in social media increasingly focus on novel interactions involving visual media, like images and short videos, in addition to text. In this work, we tackle the problem of visual sentiment analysis of social media images – specifically, the prediction of image sentiment polarity. While previous work relied on manually labeled training sets, we propose an automated approach for building sentiment polarity classifiers based on a cross-modal distillation paradigm; starting from scraped multimodal (text + images) data, we train a student model on the visual modality based on the outputs of a textual teacher model that analyses the sentiment of the corresponding textual modality. We applied our method to randomly collected images crawled from Twitter over three months and produced, after automatic cleaning, a weakly-labeled dataset of ∼1.5 million images. Despite exploiting noisy labeled samples, our training pipeline produces classifiers showing strong generalization capabilities and outperforming the current state of the art on five manually labeled benchmarks for image sentiment polarity prediction.

Abstract Unlabeled entity deduplication is a relevant task already studied in the recent literature. Most methods can be traced back to the following workflow: entity blocking phase, in-block pairwise comparisons between entities to draw similarity relations, closure of the resulting meshes to create groups of duplicate entities, and merging group entities to remove disambiguation. Such methods are effective but still not good enough whenever a very low false positive rate is required. In this paper, we present an approach for evaluating the correctness of “groups of duplicates”, which can be used to measure the group’s accuracy hence its likelihood of false-positiveness. Our novel approach is based on a Graph Neural Network that exploits and combines the concept of Graph Attention and Long Short Term Memory (LSTM). The accuracy of the proposed approach is verified in the context of Author Name Disambiguation applied to a curated dataset obtained as a subset of the OpenAIRE Graph that includes PubMed publications with at least one ORCID identifier.

VISIONE is a large-scale video retrieval system that integrates multiple search functionalities, including free text search, spatial color and object search, visual and semantic similarity search, and temporal search. The system leverages cutting-edge AI technology for visual analysis and advanced indexing techniques to ensure scalability. As demonstrated by its runner-up position in the 2023 Video Browser Showdown competition, VISIONE effectively integrates these capabilities to provide a comprehensive video retrieval solution. A system demo is available online, showcasing its capabilities on over 2300 hours of diverse video content (V3C1+V3C2 dataset) and 12 hours of highly redundant content (Marine dataset). The demo can be accessed at https://visione.isti.cnr.it/.

Due to recent advances in pose-estimation methods, human motion can be extracted from a common video in the form of 3D skeleton sequences. Despite wonderful application opportunities, effective and efficient content-based access to large volumes of such spatio-temporal skeleton data still remains a challenging problem. In this paper, we propose a novel content-based text-to-motion retrieval task, which aims at retrieving relevant motions based on a specified natural-language textual description. To define baselines for this uncharted task, we employ the BERT and CLIP language representations to encode the text modality and successful spatio-temporal models to encode the motion modality. We additionally introduce our transformer-based approach, called Motion Transformer (MoT), which employs divided space-time attention to effectively aggregate the different skeleton joints in space and time. Inspired by the recent progress in text-to-image/video matching, we experiment with two widely-adopted metric-learning loss functions. Finally, we set up a common evaluation protocol by defining qualitative metrics for assessing the quality of the retrieved motions, targeting the two recently-introduced KIT Motion-Language and HumanML3D datasets. The code for reproducing our results is available here: https://github.com/mesnico/text-to-motion-retrieval.

Falling is one of the most common causes of injury in all ages, especially in the elderly, where it is more frequent and severe. For this reason, a tool that can detect a fall in real time can be helpful in ensuring appropriate intervention and avoiding more serious damage. Some approaches available in the literature use sensors, wearable devices, or cameras with special features such as thermal or depth sensors. In this paper, we propose a Computer Vision deep-learning based approach for human fall detection based on largely available standard RGB cameras. A typical limitation of this kind of approaches is the lack of generalization to unseen environments. This is due to the error generated during human detection and, more generally, due to the unavailability of large-scale datasets that specialize in fall detection problems with different environments and fall types. In this work, we mitigate these limitations with a general-purpose object detector trained using a virtual world dataset in addition to real-world images. Through extensive experimental evaluation, we verified that by training our models on synthetic images as well, we were able to improve their ability to generalize. Code to reproduce results is available at https://github.com/lorepas/fallen-people-detection.

Features extracted from Deep Neural Networks (DNNs) have proven to be very effective in the context of Content Based Image Retrieval (CBIR). Recently, biologically inspired Hebbian learning algorithms have shown promises for DNN training. In this contribution, we study the performance of such algorithms in the development of feature extractors for CBIR tasks. Specifically, we consider a semi-supervised learning strategy in two steps: first, an unsupervised pre-training stage is performed using Hebbian learning on the image dataset; second, the network is fine-tuned using supervised Stochastic Gradient Descent (SGD) training. For the unsupervised pre-training stage, we explore the nonlinear Hebbian Principal Component Analysis (HPCA) learning rule. For the supervised fine-tuning stage, we assume sample efficiency scenarios, in which the amount of labeled samples is just a small fraction of the whole dataset. Our experimental analysis, conducted on the CIFAR10 and CIFAR100 datasets, shows that, when few labeled samples are available, our Hebbian approach provides relevant improvements compared to various alternative methods.

Image-text matching is gaining a leading role among tasks involving the joint understanding of vision and language. In literature, this task is often used as a pre-training objective to forge architectures able to jointly deal with images and texts. Nonetheless, it has a direct downstream application: cross-modal retrieval, which consists in finding images related to a given query text or vice-versa. Solving this task is of critical importance in cross-modal search engines. Many recent methods proposed effective solutions to the image-text matching problem, mostly using recent large vision-language (VL) Transformer networks. However, these models are often computationally expensive, especially at inference time. This prevents their adoption in large-scale cross-modal retrieval scenarios, where results should be provided to the user almost instantaneously. In this paper, we propose to fill in the gap between effectiveness and efficiency by proposing an ALign And DIstill Network (ALADIN). ALADIN first produces high-effective scores by aligning at fine-grained level images and texts. Then, it learns a shared embedding space – where an efficient kNN search can be performed – by distilling the relevance scores obtained from the fine-grained alignments. We obtained remarkable results on MS-COCO, showing that our method can compete with state-of-the-art VL Transformers while being almost 90 times faster. The code for reproducing our results is available at https://github.com/mesnico/ALADIN.

Although deep-learning-based solutions are pervading different application sectors, many doubts have arisen about their reliability and, above all, their security against threats that can mislead their decision mechanisms. In this work, we considered a particular kind of deep neural network, the Neural Ordinary Differential Equations (N-ODE) networks, which have shown intrinsic robustness against adversarial samples by properly tuning their tolerance parameter at test time. Their behaviour has never been investigated in image forensics tasks such as distinguishing between an original and an altered image. Following this direction, we demonstrate how tuning the tolerance parameter during the prediction phase can control and increase N-ODE’s robustness versus adversarial attacks. We performed experiments on basic image transformations used to generate tampered data, providing encouraging results in terms of adversarial rejection and preservation of the correct classification of pristine images.

Automatic people counting from images has recently drawn attention for urban monitoring in modern Smart Cities due to the ubiquity of surveillance camera networks. Current computer vision techniques rely on deep learning-based algorithms that estimate pedestrian densities in still, individual images. Only a bunch of works take advantage of temporal consistency in video sequences. In this work, we propose a spatio-temporal attentive neural network to estimate the number of pedestrians from surveillance videos. By taking advantage of the temporal correlation between consecutive frames, we lowered state-of-the-art count error by 5% and localization error by 7.5% on the widely-used FDST benchmark.

Anomalies are ubiquitous in all scientific fields and can express an unexpected event due to incomplete knowledge about the data distribution or an unknown process that suddenly comes into play and distorts the observations. Usually, due to such events’ rarity, to train deep learning (DL) models on the anomaly detection (AD) task, scientists only rely on “normal” data, i.e., nonanomalous samples. Thus, letting the neural network infer the distribution beneath the input data. In such a context, we propose a novel framework, named multilayer one-class classification (MOCCA), to train and test DL models on the AD task. Specifically, we applied our approach to autoencoders. A key novelty in our work stems from the explicit optimization of the intermediate representations for the task at hand. Indeed, differently from commonly used approaches that consider a neural network as a single computational block, i.e., using the output of the last layer only, MOCCA explicitly leverages the multilayer structure of deep architectures. Each layer’s feature space is optimized for AD during training, while in the test phase, the deep representations extracted from the trained layers are combined to detect anomalies. With MOCCA, we split the training process into two steps. First, the autoencoder is trained on the reconstruction task only. Then, we only retain the encoder tasked with minimizing the$L_{2}$distance between the output representation and a reference point, the anomaly-free training data centroid, at each considered layer. Subsequently, we combine the deep features extracted at the various trained layers of the encoder model to detect anomalies at inference time. To assess the performance of the models trained with MOCCA, we conduct extensive experiments on publicly available datasets, namely CIFAR10, MVTec AD, and ShanghaiTech. We show that our proposed method reaches comparable or superior performance to state-of-the-art approaches available in the literature. Finally, we provide a model analysis to give insights regarding the benefits of our training procedure.

Cross-modal retrieval is an important functionality in modern search engines, as it increases the user experience by allowing queries and retrieved objects to pertain to different modalities. In this paper, we focus on the image-sentence retrieval task, where the objective is to efficiently find relevant images for a given sentence (image-retrieval) or the relevant sentences for a given image (sentence-retrieval). Computer vision literature reports the best results on the image-sentence matching task using deep neural networks equipped with attention and self-attention mechanisms. They evaluate the matching performance on the retrieval task by performing sequential scans of the whole dataset. This method does not scale well with an increasing amount of images or captions. In this work, we explore different preprocessing techniques to produce sparsified deep multi-modal features extracting them from state-of-the-art deep-learning architectures for image-text matching. Our main objective is to lay down the paths for efficient indexing of complex multi-modal descriptions. We use the recently introduced TERN architecture as an image-sentence features extractor. It is designed for producing fixed-size 1024-d vectors describing whole images and sentences, as well as variable-length sets of 1024-d vectors describing the various building components of the two modalities (image regions and sentence words respectively). All these vectors are enforced by the TERN design to lie into the same common space. Our experiments show interesting preliminary results on the explored methods and suggest further experimentation in this important research direction.

Approximate Nearest Neighbor (ANN) search is a prevalent paradigm for searching intrinsically high dimensional objects in large-scale data sets. Recently, the permutation-based approach for ANN has attracted a lot of interest due to its versatility in being used in the more general class of metric spaces. In this approach, the entire database is ranked by a permutation distance to the query. Typically, permutations allow the efficient selection of a candidate set of results, but typically to achieve high recall or precision this set has to be reviewed using the original metric and data. This can lead to a sizeable percentage of the database being recalled, along with many expensive distance calculations. To reduce the number of metric computations and the number of database elements accessed, we propose here a re-ranking based on a local embedding using the nSimplex projection. The nSimplex projection produces Euclidean vectors from objects in metric spaces which possess the n-point property. The mapping is obtained from the distances to a set of reference objects, and the original metric can be lower bounded and upper bounded by the Euclidean distance of objects sharing the same set of references. Our approach is particularly advantageous for extensive databases or expensive metric function. We reuse the distances computed in the permutations in the first stage, and hence the memory footprint of the index is not increased. An extensive experimental evaluation of our approach is presented, demonstrating excellent results even on a set of hundreds of millions of objects.

We propose to address the issue of sample efficiency, in Deep Convolutional Neural Networks (DCNN), with a semi-supervised training strategy that combines Hebbian learning with gradient descent: all internal layers (both convolutional and fully connected) are pre-trained using an unsupervised approach based on Hebbian learning, and the last fully connected layer (the classification layer) is trained using Stochastic Gradient Descent (SGD). In fact, as Hebbian learning is an unsupervised learning method, its potential lies in the possibility of training the internal layers of a DCNN without labels. Only the final fully connected layer has to be trained with labeled examples. We performed experiments on various object recognition datasets, in different regimes of sample efficiency, comparing our semi-supervised (Hebbian for internal layers + SGD for the final fully connected layer) approach with end-to-end supervised backprop training, and with semi-supervised learning based on Variational Auto-Encoder (VAE). The results show that, in regimes where the number of available labeled samples is low, our semi-supervised approach outperforms the other approaches in almost all the cases.

Deep learning demonstrated major abilities in solving many kinds of different real-world problems in computer vision literature. However, they are still strained by simple reasoning tasks that humans consider easy to solve. In this work, we probe current state-of-the-art convolutional neural networks on a difficult set of tasks known as the same-different problems. All the problems require the same prerequisite to be solved correctly: understanding if two random shapes inside the same image are the same or not. With the experiments carried out in this work, we demonstrate that residual connections, and more generally the skip connections, seem to have only a marginal impact on the learning of the proposed problems. In particular, we experiment with DenseNets, and we examine the contribution of residual and recurrent connections in already tested architectures, ResNet-18, and CorNet-S respectively. Our experiments show that older feed-forward networks, AlexNet and VGG, are almost unable to learn the proposed problems, except in some specific scenarios. We show that recently introduced architectures can converge even in the cases where the important parts of their architecture are removed. We finally carry out some zero-shot generalization tests, and we discover that in these scenarios residual and recurrent connections can have a stronger impact on the overall test accuracy. On four difficult problems from the SVRT dataset, we can reach state-of-the-art results with respect to the previous approaches, obtaining super-human performances on three of the four problems.

Despite the evolution of deep-learning-based visual-textual processing systems, precise multi-modal matching remains a challenging task. In this work, we tackle the task of cross-modal retrieval through image-sentence matching based on word-region alignments, using supervision only at the global image-sentence level. Specifically, we present a novel approach called Transformer Encoder Reasoning and Alignment Network (TERAN). TERAN enforces a fine-grained match between the underlying components of images and sentences (i.e., image regions and words, respectively) to preserve the informative richness of both modalities. TERAN obtains state-of-the-art results on the image retrieval task on both MS-COCO and Flickr30k datasets. Moreover, on MS-COCO, it also outperforms current approaches on the sentence retrieval task. Focusing on scalable cross-modal information retrieval, TERAN is designed to keep the visual and textual data pipelines well separated. Cross-attention links invalidate any chance to separately extract visual and textual features needed for the online search and the offline indexing steps in large-scale retrieval systems. In this respect, TERAN merges the information from the two domains only during the final alignment phase, immediately before the loss computation. We argue that the fine-grained alignments produced by TERAN pave the way toward the research for effective and efficient methods for large-scale cross-modal information retrieval. We compare the effectiveness of our approach against relevant state-of-the-art methods. On the MS-COCO 1K test set, we obtain an improvement of 5.7% and 3.5% respectively on the image and the sentence retrieval tasks on the Recall@1 metric. The code used for the experiments is publicly available on GitHub at https://github.com/mesnico/TERAN .

In this work, we propose CBiGAN - a novel method for anomaly detection in images, where a consistency constraint is introduced as a regularization term in both the encoder and decoder of a BiGAN. Our model exhibits fairly good modeling power and reconstruction consistency capability. We evaluate the proposed method on MVTec AD - a real-world benchmark for unsupervised anomaly detection on high-resolution images - and compare against standard baselines and state-of-the-art approaches. Experiments show that the proposed method improves the performance of BiGAN formulations by a large margin and performs comparably to expensive state-of-the-art iterative methods while reducing the computational cost. We also observe that our model is particularly effective in texture-type anomaly detection, as it sets a new state of the art in this category. Our code is available at https://github.com/fabiocarrara/cbigan-ad/.

Image-text matching is an interesting and fascinating task in modern AI research. Despite the evolution of deep-learning-based image and text processing systems, multimodal matching remains a challenging problem. In this work, we consider the problem of accurate image-text matching for the task of multi-modal large-scale information retrieval. State-of-the-art results in image-text matching are achieved by inter-playing image and text features from the two different processing pipelines, usually using mutual attention mechanisms. However, this invalidates any chance to extract separate visual and textual features needed for later indexing steps in large-scale retrieval systems. In this regard, we introduce the Transformer Encoder Reasoning Network (TERN), an architecture built upon one of the modern relationship-aware self-attentive architectures, the Transformer Encoder (TE). This architecture is able to separately reason on the two different modalities and to enforce a final common abstract concept space by sharing the weights of the deeper transformer layers. Thanks to this design, the implemented network is able to produce compact and very rich visual and textual features available for the successive indexing step. Experiments are conducted on the MS-COCO dataset, and we evaluate the results using a discounted cumulative gain metric with relevance computed exploiting caption similarities, in order to assess possibly non-exact but relevant search results. We demonstrate that on this metric we are able to achieve state-of-the-art results in the image retrieval task. Our code is freely available at https://github.com/mesnico/TERN.

In the last years, content-based image retrieval largely benefited from representation extracted from deeper and more complex convolutional neural networks, which became more effective but also more computationally demanding. Despite existing hardware acceleration, query processing times may be easily saturated by deep feature extraction in high-throughput or real-time embedded scenarios, and usually, a trade-off between efficiency and effectiveness has to be accepted. In this work, we experiment with the recently proposed continuous neural networks defined by parametric ordinary differential equations, dubbed ODE-Nets, for adaptive extraction of image representations. Given the continuous evolution of the network hidden state, we propose to approximate the exact feature extraction by taking a previous "near-in-time" hidden state as features with a reduced computational cost. To understand the potential and the limits of this approach, we also evaluate an ODE-only architecture in which we minimize the number of classical layers in order to delegate most of the representation learning process --- and thus the feature extraction process --- to the continuous part of the model. Preliminary experiments on standard benchmarks show that we are able to dynamically control the trade-off between efficiency and effectiveness of feature extraction at inference-time by controlling the evolution of the continuous hidden state. Although ODE-only networks provide the best fine-grained control on the effectiveness-efficiency trade-off, we observed that mixed architectures perform better or comparably to standard residual nets in both the image classification and retrieval setups while using fewer parameters and retaining the controllability of the trade-off.

Nowadays, the possibilities offered by state-of-the-art deep neural networks allow the creation of systems capable of recognizing and indexing visual content with very high accuracy. Performance of these systems relies on the availability of high quality training sets, containing a large number of examples (e.g. million), in addition to the the machine learning tools themselves. For several applications, very good training sets can be obtained, for example, crawling (noisily) annotated images from the internet, or by analyzing user interaction (e.g.: on social networks). However, there are several applications for which high quality training sets are not easy to be obtained/created. Consider, as an example, a security scenario where one wants to automatically detect rarely occurring threatening events. In this respect, recently, researchers investigated the possibility of using a visual virtual environment, capable of artificially generating controllable and photo-realistic contents, to create training sets for applications with little available training images. We explored this idea to generate synthetic photo-realistic training sets to train classifiers to recognize the proper use of individual safety equipment (e.g.: worker protection helmets, high-visibility vests, ear protection devices) during risky human activities. Then, we performed domain adaptation to real images by using a very small image data set of real-world photographs. We show that training with the generated synthetic training set and using the domain adaptation step is an effective solution to address applications for which no training sets exist.

Developing abstract reasoning abilities in neural networks is an important goal towards the achievement of human-like performances on many tasks. As of now, some works have tackled this problem, developing ad-hoc architectures and reaching overall good generalization performances. In this work we try to understand to what extent state-of-the-art convolutional neural networks for image classification are able to deal with a challenging abstract problem, the so-called same-different task. This problem consists in understanding if two random shapes inside the same image are the same or not. A recent work demonstrated that simple convolutional neural networks are almost unable to solve this problem. We extend their work, showing that ResNet-inspired architectures are able to learn, while VGG cannot converge. In light of this, we suppose that residual connections have some important role in the learning process, while the depth of the network seems not so relevant. In addition, we carry out some targeted tests on the converged architectures to figure out to what extent they are able to generalize to never seen patterns. However, further investigation is needed in order to understand what are the architectural peculiarities and limits as far as abstract reasoning is concerned.

Neural networks are now used in many sectors of our daily life thanks to efficient solutions such instruments provide for diverse tasks. Leaving to artificial intelligence the chance to make choices on behalf of humans inevitably exposes these tools to be fraudulently attacked. In fact, adversarial examples, intentionally crafted to fool a neural network, can dangerously induce a misclassification though appearing innocuous for a human observer. On such a basis, this paper focuses on the problem of image classification and proposes an analysis to better insight what happens inside a convolutional neural network (CNN) when it evaluates an adversarial example. In particular, the activations of the internal network layers have been analyzed and exploited to design possible countermeasures to reduce CNN vulnerability. Experimental results confirm that layer activations can be adopted to detect adversarial inputs.

The integration of mobile and ubiquitous computing with deep learning methods is a promising emerging trend that aims at moving the processing task closer to the data source rather than bringing the data to a central node. The advantages of this approach range from bandwidth reduction, high scalability, to high reliability, just to name a few. In this paper, we propose a real-time deep learning approach to automatically detect and count vehicles in videos taken from a UAV (Unmanned Aerial Vehicle). Our solution relies on a convolutional neural network-based model fine-tuned to the specific domain of applications that is able to precisely localize instances of the vehicles using a regression approach, straight from image pixels to bounding box coordinates, reasoning globally about the image when making predictions and implicitly encoding contextual information. A comprehensive experimental evaluation on real-world datasets shows that our approach results in state-of-the-art performances. Furthermore, our solution achieves real-time performances by running at a speed of 4 Frames Per Second on an NVIDIA Jetson TX2 board, showing the potentiality of this approach for real-time processing in UAVs.

Content-Based Image Retrieval in large archives through the use of visual features has become a very attractive research topic in recent years. The cause of this strong impulse in this area of research is certainly to be attributed to the use of Convolutional Neural Network (CNN) activations as features and their outstanding performance. However, practically all the available image retrieval systems are implemented in main memory, limiting their applicability and preventing their usage in big-data applications. In this paper, we propose to transform CNN features into textual representations and index them with the well-known full-text retrieval engine Elasticsearch. We validate our approach on a novel CNN feature, namely Regional Maximum Activations of Convolutions. A preliminary experimental evaluation, conducted on the standard benchmark INRIA Holidays, shows the effectiveness and efficiency of the proposed approach and how it compares to state-of-the-art main-memory indexes.

In this paper, we adapt a surrogate text representation technique to develop efficient instance-level image retrieval using Regional Maximum Activations of Convolutions (R-MAC). R-MAC features have recently showed outstanding performance in visual instance retrieval. However, contrary to the activations of hidden layers adopting ReLU (Rectified Linear Unit), these features are dense. This constitutes an obstacle to the direct use of inverted indexes, which rely on sparsity of data. We propose the use of deep permutations, a recent approach for efficient evaluation of permutations, to generate surrogate text representation of R-MAC features, enabling indexing of visual features as text into a standard search-engine. The experiments, conducted on Lucene, show the effectiveness and efficiency of the proposed approach.

This paper presents an approach for real-time car parking occupancy detection that uses a Convolutional Neural Network (CNN) classifier running on-board of a smart camera with limited resources. Experiments show that our technique is very effective and robust to light condition changes, presence of shadows, and partial occlusions. The detection is reliable, even when tests are performed using images captured from a viewpoint different than the viewpoint used for training. In addition, it also demonstrates its robustness when training and tests are executed on different parking lots. We have tested and compared our solution against state of the art techniques, using a reference benchmark for parking occupancy detection. We have also produced and made publicly available an additional dataset that contains images of the parking lot taken from different viewpoints and in different days with different light conditions. The dataset captures occlusion and shadows that might disturb the classification of the parking spaces status.

Vector of locally aggregated descriptors (VLAD) is a promising approach for addressing the problem of image search on a very large scale. This representation is proposed to overcome the quantization error problem faced in Bag-of-Words (BoW) representation. In this paper, we propose to enable inverted files of standard text search engines to exploit VLAD representation to deal with large-scale image search scenarios. We show that the use of inverted files with VLAD significantly outperforms BoW in terms of efficiency and effectiveness on the same hardware and software infrastructure.

We present the VIsual Support to Interactive TOurism in Tuscany (VISITO Tuscany) project which offers an interactive guide for tourists visiting cities of art accessible via smartphones. The peculiarity of the system is that user interaction is mainly obtained by the use of images -- In order to receive information on a particular monument users just have to take a picture of it. VISITO Tuscany, using techniques of image analysis and content recognition, automatically recognize the photographed monuments and pertinent information is displayed to the user. In this paper we illustrate how the use of landmarks recognition from mobile devices can provide the tourist with relevant and customized information about various type of objects in cities of art.

Applications of image content recognition, as for instance landmark recognition, can be obtained by using techniques of kNN classifications based on the use of local image features, such as SIFT or SURF. Quality of image classification can be improved by defining geometric consistency check rules based on space transformations of the scene depicted in images. However, this prevents the use of state of the art access methods for similarity searching and sequential scan of the images in the training sets has to be executed in order to perform classification. In this paper we propose a technique that allows one to use access methods for similarity searching, such as those exploiting metric space properties, in order to perform kNN classification with geometric consistency checks. We will see that the proposed approach, in addition to offer an obvious efficiency improvement, surprisingly offers also an improvement of the effectiveness of the classification.

In order to become an effective complement to traditional Web-scale text-based image retrieval solutions, content-based image retrieval must address scalability and efficiency issues. In this paper we investigate the possibility of caching the answers to content-based image retrieval queries in metric space, with the aim of reducing the average cost of query processing, and boosting the overall system throughput. Our proposal exploits the similarity between the query object and the cache content, and allows the cache to return approximate answers with acceptable quality guarantee even if the query processed has never been encountered in the past. Moreover, since popular images that are likely to be used as query have several near-duplicate versions, we show that our caching algorithm is robust, and does not suffer of cache pollution problems due to near-duplicate query objects. We report on very promising results obtained with a collection of one million high-quality digital photos. We show that it is worth pursuing caching strategies also in similarity search systems, since the proposed caching techniques can have a significant impact on performance, like caching on text queries has been proven effective for traditional Web search engines.