Egor Bondarev

Giacomo D'Amicantonio, Egor Bondarev, Peter H. N. De With

Over the past few decades, a significant rise of camera-based applications for traffic monitoring has occurred. Governments and local administrations are increasingly relying on the data collected from these cameras to enhance road safety and optimize traffic conditions. However, for effective data utilization, it is imperative to ensure accurate and automated calibration of the involved cameras. This paper proposes a novel approach to address this challenge by leveraging the topological structure of intersections. We propose a framework involving the generation of a set of synthetic intersection viewpoint images from a bird's-eye-view image, framed as a graph of virtual cameras to model these images. Using the capabilities of Graph Neural Networks, we effectively learn the relationships within this graph, thereby facilitating the estimation of a homography matrix. This estimation leverages the neighbourhood representation for any real-world camera and is enhanced by exploiting multiple images instead of a single match. In turn, the homography matrix allows the retrieval of extrinsic calibration parameters. As a result, the proposed framework demonstrates superior performance on both synthetic datasets and real-world cameras, setting a new state-of-the-art benchmark.

Camile Lendering, Erkut Akdag, Egor Bondarev

Accurate defect segmentation is critical for industrial inspection, yet dense pixel-level annotations are rarely available. A common workaround is to convert inexpensive bounding boxes into pseudo-masks using foundation segmentation models such as the Segment Anything Model (SAM). However, these pseudo-labels are systematically noisy on industrial surfaces, often hallucinating background structure while missing sparse defects. To address this limitation, a noise-robust box-to-pixel distillation framework, Boxes2Pixels, is proposed that treats SAM as a noisy teacher rather than a source of ground-truth supervision. Bounding boxes are converted into pseudo-masks offline by SAM, and a compact student is trained with (i) a hierarchical decoder over frozen DINOv2 features for semantic stability, (ii) an auxiliary binary localization head to decouple sparse foreground discovery from class prediction, and (iii) a one-sided online self-correction mechanism that relaxes background supervision when the student is confident, targeting teacher false negatives. On a manually annotated wind turbine inspection benchmark, the proposed Boxes2Pixels improves anomaly mIoU by +6.97 and binary IoU by +9.71 over the strongest baseline trained under identical weak supervision. Moreover, online self-correction increases the binary recall by +18.56, while the model employs 80\% fewer trainable parameters. Code is available at https://github.com/CLendering/Boxes2Pixels.

Pedro Quesado, Erkut Akdag, Yasaman Kashefbahrami et al.

Live-streaming Novel View Synthesis (NVS) from unposed multi-view video remains an open challenge in a wide range of applications. Existing methods for dynamic scene representation typically require ground-truth camera parameters and involve lengthy optimizations ($\approx 2.67$s), which makes them unsuitable for live streaming scenarios. To address this issue, we propose a novel viewpoint video live-streaming method (LiveStre4m), a feed-forward model for real-time NVS from unposed sparse multi-view inputs. LiveStre4m introduces a multi-view vision transformer for keyframe 3D scene reconstruction coupled with a diffusion-transformer interpolation module that ensures temporal consistency and stable streaming. In addition, a Camera Pose Predictor module is proposed to efficiently estimate both poses and intrinsics directly from RGB images, removing the reliance on known camera calibration information. Our approach enables temporally consistent novel-view video streaming in real-time using as few as two synchronized unposed input streams. LiveStre4m attains an average reconstruction time of $ 0.07$s per-frame at $ 1024 \times 768$ resolution, outperforming the optimization-based dynamic scene representation methods by orders of magnitude in runtime. These results demonstrate that LiveStre4m makes real-time NVS streaming feasible in practical settings, marking a substantial step toward deployable live novel-view synthesis systems. Code available at: https://github.com/pedro-quesado/LiveStre4m

Camile Lendering, Erkut Akdag, Egor Bondarev

Detecting visual anomalies in industrial inspection often requires training with only a few normal images per category. Recent few-shot methods achieve strong results employing foundation-model features, but typically rely on memory banks, auxiliary datasets, or multi-modal tuning of vision-language models. We therefore question whether such complexity is necessary given the feature representations of vision foundation models. To answer this question, we introduce SubspaceAD, a training-free method, that operates in two simple stages. First, patch-level features are extracted from a small set of normal images by a frozen DINOv2 backbone. Second, a Principal Component Analysis (PCA) model is fit to these features to estimate the low-dimensional subspace of normal variations. At inference, anomalies are detected via the reconstruction residual with respect to this subspace, producing interpretable and statistically grounded anomaly scores. Despite its simplicity, SubspaceAD achieves state-of-the-art performance across one-shot and few-shot settings without training, prompt tuning, or memory banks. In the one-shot anomaly detection setting, SubspaceAD achieves image-level and pixel-level AUROC of 98.0% and 97.6% on the MVTec-AD dataset, and 93.3% and 98.3% on the VisA dataset, respectively, surpassing prior state-of-the-art results. Code and demo are available at https://github.com/CLendering/SubspaceAD.

Joaquín Figueira, Rob Van Gastel, Giacomo D'Amicantonio et al.

Segmentation models in automated optical inspection of wire-bonded semiconductors are typically device-specific and must be re-trained when new devices or distribution shifts appear. We introduce AOI-SSL, a training-efficient framework for semantic segmentation of wire-bonded semiconductors by combining small-domain self-supervised pre-training of vision transformers with in-context inference that minimizes the need of labeled examples. We pre-train SOTA self-supervised algorithms in a small industrial inspection dataset and find that Masked Autoencoders are the most effective in this small-data setting, improving downstream segmentation while reducing the labeled fine-tuning effort. We further introduce in-context, patch-level retrieval methods that predict masks directly from dense encoder embeddings with negligible additional training. We show that, in this setting, simple similarity-based retrieval performs on par with more complex attention-based aggregation used currently in the literature. Furthermore, our experiments demonstrate that self-supervised pre-training significantly improves segmentation quality compared to training from scratch and to ImageNet pre-trained backbones under a fixed fine-tuning computational budget. Finally, the results reveal that retrieval based segmentation outperforms fine-tuning when targeting single device images, allowing for near-instant adaptation to difficult samples.

Willem Menu, Erkut Akdag, Pedro Quesado et al.

Existing full-body Gaussian avatar methods primarily optimize global reconstruction quality and often fail to preserve fine-grained facial geometry and expression details. This challenge arises from limited facial representational capacity that causes difficulties in modeling high-frequency pose-dependent deformations. To address this, we propose F3G-Avatar, a full-body, face-aware avatar synthesis method that reconstructs animatable human representations from multi-view RGB video and regressed pose/shape parameters. Starting from a clothed Momentum Human Rig (MHR) template, front/back positional maps are rendered and decoded into 3D Gaussians through a two-branch architecture: a body branch that captures pose-dependent non-rigid deformations and a face-focused deformation branch that refines head geometry and appearance. The predicted Gaussians are fused, posed with linear blend skinning (LBS), and rendered with differentiable Gaussian splatting. Training combines reconstruction and perceptual objectives with a face-specific adversarial loss to enhance realism in close-up views. Experiments demonstrate strong rendering quality, with face-view performance reaching PSNR/SSIM/LPIPS of 26.243/0.964/0.084 on the AvatarReX dataset. Ablations further highlight contributions of the MHR template and the face-focused deformation. F3G-Avatar provides a practical, high-quality pipeline for realistic, animatable full-body avatar synthesis.

Erkut Akdag, Zeqi Zhu, Egor Bondarev et al.

Classification and localization of driving actions over time is important for advanced driver-assistance systems and naturalistic driving studies. Temporal localization is challenging because it requires robustness, reliability, and accuracy. In this study, we aim to improve the temporal localization and classification accuracy performance by adapting video action recognition and 2D human-pose estimation networks to one model. Therefore, we design a transformer-based fusion architecture to effectively combine 2D-pose features and spatio-temporal features. The model uses 2D-pose features as the positional embedding of the transformer architecture and spatio-temporal features as the main input to the encoder of the transformer. The proposed solution is generic and independent of the camera numbers and positions, giving frame-based class probabilities as output. Finally, the post-processing step combines information from different camera views to obtain final predictions and eliminate false positives. The model performs well on the A2 test set of the 2023 NVIDIA AI City Challenge for naturalistic driving action recognition, achieving the overlap score of the organizer-defined distracted driver behaviour metric of 0.5079.

Tunc Alkanat, Erkut Akdag, Egor Bondarev et al.

Temporal localization of driving actions plays a crucial role in advanced driver-assistance systems and naturalistic driving studies. However, this is a challenging task due to strict requirements for robustness, reliability and accurate localization. In this work, we focus on improving the overall performance by efficiently utilizing video action recognition networks and adapting these to the problem of action localization. To this end, we first develop a density-guided label smoothing technique based on label probability distributions to facilitate better learning from boundary video-segments that typically include multiple labels. Second, we design a post-processing step to efficiently fuse information from video-segments and multiple camera views into scene-level predictions, which facilitates elimination of false positives. Our methodology yields a competitive performance on the A2 test set of the naturalistic driving action recognition track of the 2022 NVIDIA AI City Challenge with an F1 score of 0.271.

Ivo P. C. Kersten, Erkut Akdag, Egor Bondarev et al.

Anomalous behavior detection is a challenging research area within computer vision. Progress in this area enables automated detection of dangerous behavior using surveillance camera feeds. A dangerous behavior that is often overlooked in other research is the throwing action in traffic flow, which is one of the unique requirements of our Smart City project to enhance public safety. This paper proposes a solution for throwing action detection in surveillance videos using deep learning. At present, datasets for throwing actions are not publicly available. To address the use-case of our Smart City project, we first generate the novel public 'Throwing Action' dataset, consisting of 271 videos of throwing actions performed by traffic participants, such as pedestrians, bicyclists, and car drivers, and 130 normal videos without throwing actions. Second, we compare the performance of different feature extractors for our anomaly detection method on the UCF-Crime and Throwing-Action datasets. The explored feature extractors are the Convolutional 3D (C3D) network, the Inflated 3D ConvNet (I3D) network, and the Multi-Fiber Network (MFNet). Finally, the performance of the anomaly detection algorithm is improved by applying the Adam optimizer instead of Adadelta, and proposing a mean normal loss function that covers the multitude of normal situations in traffic. Both aspects yield better anomaly detection performance. Besides this, the proposed mean normal loss function lowers the false alarm rate on the combined dataset. The experimental results reach an area under the ROC curve of 86.10 for the Throwing-Action dataset, and 80.13 on the combined dataset, respectively.

Erkut Akdag, Egor Bondarev, Peter H. N. De With

Anomaly detection in video surveillance has recently gained interest from the research community. Temporal duration of anomalies vary within video streams, leading to complications in learning the temporal dynamics of specific events. This paper presents a temporal-granularity method for an anomaly detection model (TeG) in real-world surveillance, combining spatio-temporal features at different time-scales. The TeG model employs multi-head cross-attention blocks and multi-head self-attention blocks for this purpose. Additionally, we extend the UCF-Crime dataset with new anomaly types relevant to Smart City research project. The TeG model is deployed and validated in a city surveillance system, achieving successful real-time results in industrial settings.

Giacomo D'Amicantonio, Snehashis Majhi, Quan Kong et al.

Video Anomaly Detection (VAD) is a challenging task due to the variability of anomalous events and the limited availability of labeled data. Under the Weakly-Supervised VAD (WSVAD) paradigm, only video-level labels are provided during training, while predictions are made at the frame level. Although state-of-the-art models perform well on simple anomalies (e.g., explosions), they struggle with complex real-world events (e.g., shoplifting). This difficulty stems from two key issues: (1) the inability of current models to address the diversity of anomaly types, as they process all categories with a shared model, overlooking category-specific features; and (2) the weak supervision signal, which lacks precise temporal information, limiting the ability to capture nuanced anomalous patterns blended with normal events. To address these challenges, we propose Gaussian Splatting-guided Mixture of Experts (GS-MoE), a novel framework that employs a set of expert models, each specialized in capturing specific anomaly types. These experts are guided by a temporal Gaussian splatting loss, enabling the model to leverage temporal consistency and enhance weak supervision. The Gaussian splatting approach encourages a more precise and comprehensive representation of anomalies by focusing on temporal segments most likely to contain abnormal events. The predictions from these specialized experts are integrated through a mixture-of-experts mechanism to model complex relationships across diverse anomaly patterns. Our approach achieves state-of-the-art performance, with a 91.58% AUC on the UCF-Crime dataset, and demonstrates superior results on XD-Violence and MSAD datasets. By leveraging category-specific expertise and temporal guidance, GS-MoE sets a new benchmark for VAD under weak supervision.

Snehashis Majhi, Giacomo D'Amicantonio, Antitza Dantcheva et al.

Weakly-supervised methods for video anomaly detection (VAD) are conventionally based merely on RGB spatio-temporal features, which continues to limit their reliability in real-world scenarios. This is due to the fact that RGB-features are not sufficiently distinctive in setting apart categories such as shoplifting from visually similar events. Therefore, towards robust complex real-world VAD, it is essential to augment RGB spatio-temporal features by additional modalities. Motivated by this, we introduce the Poly-modal Induced framework for VAD: "PI-VAD", a novel approach that augments RGB representations by five additional modalities. Specifically, the modalities include sensitivity to fine-grained motion (Pose), three dimensional scene and entity representation (Depth), surrounding objects (Panoptic masks), global motion (optical flow), as well as language cues (VLM). Each modality represents an axis of a polygon, streamlined to add salient cues to RGB. PI-VAD includes two plug-in modules, namely Pseudo-modality Generation module and Cross Modal Induction module, which generate modality-specific prototypical representation and, thereby, induce multi-modal information into RGB cues. These modules operate by performing anomaly-aware auxiliary tasks and necessitate five modality backbones -- only during training. Notably, PI-VAD achieves state-of-the-art accuracy on three prominent VAD datasets encompassing real-world scenarios, without requiring the computational overhead of five modality backbones at inference.

Cris Claessens, Christiaan Viviers, Giacomo D'Amicantonio et al.

We introduce SPECTRE, a fully transformer-based foundation model for volumetric computed tomography (CT). Our Self-Supervised & Cross-Modal Pretraining for CT Representation Extraction (SPECTRE) approach utilizes scalable 3D Vision Transformer architectures and modern self-supervised and vision-language pretraining strategies to learn general-purpose CT representations. Volumetric CT poses unique challenges, such as extreme token scaling, geometric anisotropy, and weak or noisy clinical supervision, that make standard transformer and contrastive learning recipes ineffective out of the box. The framework jointly optimizes a local transformer for high-resolution volumetric feature extraction and a global transformer for whole-scan context modeling, making large-scale 3D attention computationally tractable. Notably, SPECTRE is trained exclusively on openly available CT datasets, demonstrating that high-performing, generalizable representations can be achieved without relying on private data. Pretraining combines DINO-style self-distillation with SigLIP-based vision-language alignment using paired radiology reports, yielding features that are both geometrically consistent and clinically meaningful. Across multiple CT benchmarks, SPECTRE consistently outperforms prior CT foundation models in both zero-shot and fine-tuned settings, establishing SPECTRE as a scalable, open, and fully transformer-based foundation model for 3D medical imaging.

Sara Abdulaziz, Giacomo D'Amicantonio, Egor Bondarev

Recent advances in AI-powered surveillance have intensified concerns over the collection and processing of sensitive personal data. In response, research has increasingly focused on privacy-by-design solutions, raising the need for objective techniques to evaluate privacy protection. This paper presents a comprehensive framework for evaluating visual privacy-protection methods across three dimensions: privacy, utility, and practicality. In addition, it introduces HR-VISPR, a publicly available human-centric dataset with biometric, soft-biometric, and non-biometric labels to train an interpretable privacy metric. We evaluate 11 privacy protection methods, ranging from conventional techniques to advanced deep-learning methods, through the proposed framework. The framework differentiates privacy levels in alignment with human visual perception, while highlighting trade-offs between privacy, utility, and practicality. This study, along with the HR-VISPR dataset, serves as an insightful tool and offers a structured evaluation framework applicable across diverse contexts.

Sara Abdulaziz, Egor Bondarev

Advancements in deep learning have improved anomaly detection in surveillance videos, yet they raise urgent privacy concerns due to the collection of sensitive human data. In this paper, we present a comprehensive analysis of anomaly detection performance under four human anonymization techniques, including blurring, masking, encryption, and avatar replacement, applied to the UCF-Crime dataset. We evaluate four anomaly detection methods, MGFN, UR-DMU, BN-WVAD, and PEL4VAD, on the anonymized UCF-Crime to reveal how each method responds to different obfuscation techniques. Experimental results demonstrate that anomaly detection remains viable under anonymized data and is dependent on the algorithmic design and the learning strategy. For instance, under certain anonymization patterns, such as encryption and masking, some models inadvertently achieve higher AUC performance compared to raw data, due to the strong responsiveness of their algorithmic components to these noise patterns. These results highlight the algorithm-specific sensitivities to anonymization and emphasize the trade-off between preserving privacy and maintaining detection utility. Furthermore, we compare these conventional anonymization techniques with the emerging privacy-by-design solutions, highlighting an often overlooked trade-off between robust privacy protection and utility flexibility. Through comprehensive experiments and analyses, this study provides a compelling benchmark and insights into balancing human privacy with the demands of anomaly detection.

Clint Sebastian, Raffaele Imbriaco, Egor Bondarev et al.

Vehicle re-identification plays a crucial role in the management of transportation infrastructure and traffic flow. However, this is a challenging task due to the large view-point variations in appearance, environmental and instance-related factors. Modern systems deploy CNNs to produce unique representations from the images of each vehicle instance. Most work focuses on leveraging new losses and network architectures to improve the descriptiveness of these representations. In contrast, our work concentrates on re-ranking and embedding expansion techniques. We propose an efficient approach for combining the outputs of multiple models at various scales while exploiting tracklet and neighbor information, called dual embedding expansion (DEx). Additionally, a comparative study of several common image retrieval techniques is presented in the context of vehicle re-ID. Our system yields competitive performance in the 2020 NVIDIA AI City Challenge with promising results. We demonstrate that DEx when combined with other re-ranking techniques, can produce an even larger gain without any additional attribute labels or manual supervision.

Clint Sebastian, Raffaele Imbriaco, Egor Bondarev et al.

Building extraction from aerial images has several applications in problems such as urban planning, change detection, and disaster management. With the increasing availability of data, Convolutional Neural Networks (CNNs) for semantic segmentation of remote sensing imagery has improved significantly in recent years. However, convolutions operate in local neighborhoods and fail to capture non-local features that are essential in semantic understanding of aerial images. In this work, we propose to improve building segmentation of different sizes by capturing long-range dependencies using contextual pyramid attention (CPA). The pathways process the input at multiple scales efficiently and combine them in a weighted manner, similar to an ensemble model. The proposed method obtains state-of-the-art performance on the Inria Aerial Image Labelling Dataset with minimal computation costs. Our method improves 1.8 points over current state-of-the-art methods and 12.6 points higher than existing baselines on the Intersection over Union (IoU) metric without any post-processing. Code and models will be made publicly available.

Clint Sebastian, Raffaele Imbriaco, Egor Bondarev et al.

Automatic building extraction from aerial imagery has several applications in urban planning, disaster management, and change detection. In recent years, several works have adopted deep convolutional neural networks (CNNs) for building extraction, since they produce rich features that are invariant against lighting conditions, shadows, etc. Although several advances have been made, building extraction from aerial imagery still presents multiple challenges. Most of the deep learning segmentation methods optimize the per-pixel loss with respect to the ground truth without knowledge of the context. This often leads to imperfect outputs that may lead to missing or unrefined regions. In this work, we propose a novel loss function combining both adversarial and cross-entropy losses that learn to understand both local and global contexts for semantic segmentation. The newly proposed loss function deployed on the DeepLab v3+ network obtains state-of-the-art results on the Massachusetts buildings dataset. The loss function improves the structure and refines the edges of buildings without requiring any of the commonly used post-processing methods, such as Conditional Random Fields. We also perform ablation studies to understand the impact of the adversarial loss. Finally, the proposed method achieves a relaxed F1 score of 95.59% on the Massachusetts buildings dataset compared to the previous best F1 of 94.88%.

Raffaele Imbriaco, Clint Sebastian, Egor Bondarev et al.

Remote Sensing Image Retrieval remains a challenging topic due to the special nature of Remote Sensing Imagery. Such images contain various different semantic objects, which clearly complicates the retrieval task. In this paper, we present an image retrieval pipeline that uses attentive, local convolutional features and aggregates them using the Vector of Locally Aggregated Descriptors (VLAD) to produce a global descriptor. We study various system parameters such as the multiplicative and additive attention mechanisms and descriptor dimensionality. We propose a query expansion method that requires no external inputs. Experiments demonstrate that even without training, the local convolutional features and global representation outperform other systems. After system tuning, we can achieve state-of-the-art or competitive results. Furthermore, we observe that our query expansion method increases overall system performance by about 3%, using only the top-three retrieved images. Finally, we show how dimensionality reduction produces compact descriptors with increased retrieval performance and fast retrieval computation times, e.g. 50% faster than the current systems.

Clint Sebastian, Bas Boom, Egor Bondarev et al.

Recently, privacy has a growing importance in several domains, especially in street-view images. The conventional way to achieve this is to automatically detect and blur sensitive information from these images. However, the processing cost of blurring increases with the ever-growing resolution of images. We propose a system that is cost-effective even after increasing the resolution by a factor of 2.5. The new system utilizes depth data obtained from LiDAR to significantly reduce the search space for detection, thereby reducing the processing cost. Besides this, we test several detectors after reducing the detection space and provide an alternative solution based on state-of-the-art deep learning detectors to the existing HoG-SVM-Deep system that is faster and has a higher performance.

Raffaele Imbriaco, Clint Sebastian, Egor Bondarev et al.

In this work, we present a camera geopositioning system based on matching a query image against a database with panoramic images. For matching, our system uses memory vectors aggregated from global image descriptors based on convolutional features to facilitate fast searching in the database. To speed up searching, a clustering algorithm is used to balance geographical positioning and computation time. We refine the obtained position from the query image using a new outlier removal algorithm. The matching of the query image is obtained with a recall@5 larger than 90% for panorama-to-panorama matching. We cluster available panoramas from geographically adjacent locations into a single compact representation and observe computational gains of approximately 50% at the cost of only a small (approximately 3%) recall loss. Finally, we present a coordinate estimation algorithm that reduces the median geopositioning error by up to 20%.

Clint Sebastian, Bas Boom, Thijs van Lankveld et al.

Detection of buildings and other objects from aerial images has various applications in urban planning and map making. Automated building detection from aerial imagery is a challenging task, as it is prone to varying lighting conditions, shadows and occlusions. Convolutional Neural Networks (CNNs) are robust against some of these variations, although they fail to distinguish easy and difficult examples. We train a detection algorithm from RGB-D images to obtain a segmented mask by using the CNN architecture DenseNet.First, we improve the performance of the model by applying a statistical re-sampling technique called Bootstrapping and demonstrate that more informative examples are retained. Second, the proposed method outperforms the non-bootstrapped version by utilizing only one-sixth of the original training data and it obtains a precision-recall break-even of 95.10% on our aerial imagery dataset.