Jan van Gemert

Ruben Sangers, Jan van Gemert, Sander van Cranenburgh

Deep learning based computer vision models are increasingly used by urban planners to support decision making for shaping urban environments. Such models predict how people perceive the urban environment quality in terms of e.g. its safety or beauty. However, the blackbox nature of deep learning models hampers urban planners to understand what landscape objects contribute to a particularly high quality or low quality urban space perception. This study investigates how computer vision models can be used to extract relevant policy information about peoples' perception of the urban space. To do so, we train two widely used computer vision architectures; a Convolutional Neural Network and a transformer, and apply GradCAM -- a well-known ex-post explainable AI technique -- to highlight the image regions important for the model's prediction. Using these GradCAM visualizations, we manually annotate the objects relevant to the models' perception predictions. As a result, we are able to discover new objects that are not represented in present object detection models used for annotation in previous studies. Moreover, our methodological results suggest that transformer architectures are better suited to be used in combination with GradCAM techniques. Code is available on Github.

Joris Quist, Yunqiang Li, Jan van Gemert

Binary Neural Networks (BNNs) are compact and efficient by using binary weights instead of real-valued weights. Current BNNs use latent real-valued weights during training, where several training hyper-parameters are inherited from real-valued networks. The interpretation of several of these hyperparameters is based on the magnitude of the real-valued weights. For BNNs, however, the magnitude of binary weights is not meaningful, and thus it is unclear what these hyperparameters actually do. One example is weight-decay, which aims to keep the magnitude of real-valued weights small. Other examples are latent weight initialization, the learning rate, and learning rate decay, which influence the magnitude of the real-valued weights. The magnitude is interpretable for real-valued weights, but loses its meaning for binary weights. In this paper we offer a new interpretation of these magnitude-based hyperparameters based on higher-order gradient filtering during network optimization. Our analysis makes it possible to understand how magnitude-based hyperparameters influence the training of binary networks which allows for new optimization filters specifically designed for binary neural networks that are independent of their real-valued interpretation. Moreover, our improved understanding reduces the number of hyperparameters, which in turn eases the hyperparameter tuning effort which may lead to better hyperparameter values for improved accuracy. Code is available at https://github.com/jorisquist/Understanding-WM-HP-in-BNNs

Liang Zeng, Attila Lengyel, Nergis Tömen et al.

In this work, we leverage estimated depth to boost self-supervised contrastive learning for segmentation of urban scenes, where unlabeled videos are readily available for training self-supervised depth estimation. We argue that the semantics of a coherent group of pixels in 3D space is self-contained and invariant to the contexts in which they appear. We group coherent, semantically related pixels into coherent depth regions given their estimated depth and use copy-paste to synthetically vary their contexts. In this way, cross-context correspondences are built in contrastive learning and a context-invariant representation is learned. For unsupervised semantic segmentation of urban scenes, our method surpasses the previous state-of-the-art baseline by +7.14% in mIoU on Cityscapes and +6.65% on KITTI. For fine-tuning on Cityscapes and KITTI segmentation, our method is competitive with existing models, yet, we do not need to pre-train on ImageNet or COCO, and we are also more computationally efficient. Our code is available on https://github.com/LeungTsang/CPCDR

Marian Bittner, Wei-Tse Yang, Xucong Zhang et al.

Markerless estimation of 3D Kinematics has the great potential to clinically diagnose and monitor movement disorders without referrals to expensive motion capture labs; however, current approaches are limited by performing multiple de-coupled steps to estimate the kinematics of a person from videos. Most current techniques work in a multi-step approach by first detecting the pose of the body and then fitting a musculoskeletal model to the data for accurate kinematic estimation. Errors in training data of the pose detection algorithms, model scaling, as well the requirement of multiple cameras limit the use of these techniques in a clinical setting. Our goal is to pave the way toward fast, easily applicable and accurate 3D kinematic estimation \xdeleted{in a clinical setting}. To this end, we propose a novel approach for direct 3D human kinematic estimation D3KE from videos using deep neural networks. Our experiments demonstrate that the proposed end-to-end training is robust and outperforms 2D and 3D markerless motion capture based kinematic estimation pipelines in terms of joint angles error by a large margin (35\% from 5.44 to 3.54 degrees). We show that D3KE is superior to the multi-step approach and can run at video framerate speeds. This technology shows the potential for clinical analysis from mobile devices in the future.

Yeshwanth Napolean, Anwesh Marwade, Nergis Tomen et al.

Estimating heart rate from video allows non-contact health monitoring with applications in patient care, human interaction, and sports. Existing work can robustly measure heart rate under some degree of motion by face tracking. However, this is not always possible in unconstrained settings, as the face might be occluded or even outside the camera. Here, we present IntensePhysio: a challenging video heart rate estimation dataset with realistic face occlusions, severe subject motion, and ample heart rate variation. To ensure heart rate variation in a realistic setting we record each subject for around 1-2 hours. The subject is exercising (at a moderate to high intensity) on a cycling ergometer with an attached video camera and is given no instructions regarding positioning or movement. We have 11 subjects, and approximately 20 total hours of video. We show that the existing remote photo-plethysmography methods have difficulty in estimating heart rate in this setting. In addition, we present IBIS-CNN, a new baseline using spatio-temporal superpixels, which improves on existing models by eliminating the need for a visible face/face tracking. We will make the code and data publically available soon.

Robert-Jan Bruintjes, Tomasz Motyka, Jan van Gemert

Equivariance w.r.t. geometric transformations in neural networks improves data efficiency, parameter efficiency and robustness to out-of-domain perspective shifts. When equivariance is not designed into a neural network, the network can still learn equivariant functions from the data. We quantify this learned equivariance, by proposing an improved measure for equivariance. We find evidence for a correlation between learned translation equivariance and validation accuracy on ImageNet. We therefore investigate what can increase the learned equivariance in neural networks, and find that data augmentation, reduced model capacity and inductive bias in the form of convolutions induce higher learned equivariance in neural networks.

Ombretta Strafforello, Klamer Schutte, Jan van Gemert

Many real-world applications, from sport analysis to surveillance, benefit from automatic long-term action recognition. In the current deep learning paradigm for automatic action recognition, it is imperative that models are trained and tested on datasets and tasks that evaluate if such models actually learn and reason over long-term information. In this work, we propose a method to evaluate how suitable a video dataset is to evaluate models for long-term action recognition. To this end, we define a long-term action as excluding all the videos that can be correctly recognized using solely short-term information. We test this definition on existing long-term classification tasks on three popular real-world datasets, namely Breakfast, CrossTask and LVU, to determine if these datasets are truly evaluating long-term recognition. Our study reveals that these datasets can be effectively solved using shortcuts based on short-term information. Following this finding, we encourage long-term action recognition researchers to make use of datasets that need long-term information to be solved.

Athanasios Masouris, Jan van Gemert

Chess recognition is the task of extracting the chess piece configuration from a chessboard image. Current approaches use a pipeline of separate, independent, modules such as chessboard detection, square localization, and piece classification. Instead, we follow the deep learning philosophy and explore an end-to-end approach to directly predict the configuration from the image, thus avoiding the error accumulation of the sequential approaches and eliminating the need for intermediate annotations. Furthermore, we introduce a new dataset, Chess Recognition Dataset (ChessReD), that consists of 10,800 real photographs and their corresponding annotations. In contrast to existing datasets that are synthetically rendered and have only limited angles, ChessReD has photographs captured from various angles using smartphone cameras; a sensor choice made to ensure real-world applicability. Our approach in chess recognition on the introduced challenging benchmark dataset outperforms related approaches, successfully recognizing the chess pieces' configuration in 15.26% of ChessReD's test images. This accuracy may seem low, but it is ~7x better than the current state-of-the-art and reflects the difficulty of the problem. The code and data are available through: https://github.com/ThanosM97/end-to-end-chess-recognition.

Thomas Markhorst, Zhi-Yi Lin, Jouh Yeong Chew et al.

Multi-person social interactions are inherently built on coherence and relationships among all individuals within the group, making multi-person localization and body pose estimation essential to understanding these social dynamics. One promising approach is 2D-to-3D pose lifting which provides a 3D human pose consisting of rich spatial details by building on the significant advances in 2D pose estimation. However, the existing 2D-to-3D pose lifting methods often neglect inter-person relationships or cannot handle varying group sizes, limiting their effectiveness in multi-person settings. We propose MuPPet, a novel multi-person 2D-to-3D pose lifting framework that explicitly models inter-person correlations. To leverage these inter-person dependencies, our approach introduces Person Encoding to structure individual representations, Permutation Augmentation to enhance training diversity, and Dynamic Multi-Person Attention to adaptively model correlations between individuals. Extensive experiments on group interaction datasets demonstrate MuPPet significantly outperforms state-of-the-art single- and multi-person 2D-to-3D pose lifting methods, and improves robustness in occlusion scenarios. Our findings highlight the importance of modeling inter-person correlations, paving the way for accurate and socially-aware 3D pose estimation. Our code is available at: https://github.com/Thomas-Markhorst/MuPPet

Ana Baltaretu, Pascal Benschop, Jan van Gemert

Human Action Recognition (HAR) models are increasingly deployed in high-stakes environments, yet their fairness across different human appearances has not been analyzed. We introduce a framework for auditing bias in HAR models using synthetic video data, generated with full control over visual identity attributes such as skin color. Unlike prior work that focuses on static images or pose estimation, our approach preserves temporal consistency, allowing us to isolate and test how changes to a single attribute affect model predictions. Through controlled interventions using the BEDLAM simulation platform, we show whether some popular HAR models exhibit statistically significant biases on the skin color even when the motion remains identical. Our results highlight how models may encode unwanted visual associations, and we provide evidence of systematic errors across groups. This work contributes a framework for auditing HAR models and supports the development of more transparent, accountable systems in light of upcoming regulatory standards.

Tom Edixhoven, Attila Lengyel, Jan van Gemert

In this paper we show how Group Equivariant Convolutional Neural Networks use subsampling to learn to break equivariance to their symmetries. We focus on 2D rotations and reflections and investigate the impact of broken equivariance on network performance. We show that a change in the input dimension of a network as small as a single pixel can be enough for commonly used architectures to become approximately equivariant, rather than exactly. We investigate the impact of networks not being exactly equivariant and find that approximately equivariant networks generalise significantly worse to unseen symmetries compared to their exactly equivariant counterparts. However, when the symmetries in the training data are not identical to the symmetries of the network, we find that approximately equivariant networks are able to relax their own equivariant constraints, causing them to match or outperform exactly equivariant networks on common benchmark datasets.

Sieger Falkena, Hadi Jamali-Rad, Jan van Gemert

Binary Neural Networks (BNNs) are receiving an upsurge of attention for bringing power-hungry deep learning towards edge devices. The traditional wisdom in this space is to employ sign() for binarizing featuremaps. We argue and illustrate that sign() is a uniqueness bottleneck, limiting information propagation throughout the network. To alleviate this, we propose to dispense sign(), replacing it with a learnable activation binarizer (LAB), allowing the network to learn a fine-grained binarization kernel per layer - as opposed to global thresholding. LAB is a novel universal module that can seamlessly be integrated into existing architectures. To confirm this, we plug it into four seminal BNNs and show a considerable performance boost at the cost of tolerable increase in delay and complexity. Finally, we build an end-to-end BNN (coined as LAB-BNN) around LAB, and demonstrate that it achieves competitive performance on par with the state-of-the-art on ImageNet.

Ombretta Strafforello, Vanathi Rajasekart, Osman S. Kayhan et al.

The localization quality of automatic object detectors is typically evaluated by the Intersection over Union (IoU) score. In this work, we show that humans have a different view on localization quality. To evaluate this, we conduct a survey with more than 70 participants. Results show that for localization errors with the exact same IoU score, humans might not consider that these errors are equal, and express a preference. Our work is the first to evaluate IoU with humans and makes it clear that relying on IoU scores alone to evaluate localization errors might not be sufficient.

Amogh Gudi, Fritjof Büttner, Jan van Gemert

Mean squared error (MSE) is one of the most widely used metrics to expression differences between multi-dimensional entities, including images. However, MSE is not locally sensitive as it does not take into account the spatial arrangement of the (pixel) differences, which matters for structured data types like images. Such spatial arrangements carry information about the source of the differences; therefore, an error function that also incorporates the location of errors can lead to a more meaningful distance measure. We introduce Proximally Sensitive Error (PSE), through which we suggest that a regional emphasis in the error measure can 'highlight' semantic differences between images over syntactic/random deviations. We demonstrate that this emphasis can be leveraged upon for the task of anomaly/occlusion detection. We further explore its utility as a loss function to help a model focus on learning representations of semantic objects instead of minimizing syntactic reconstruction noise.

Ombretta Strafforello, Xin Liu, Klamer Schutte et al.

Previous work on long-term video action recognition relies on deep 3D-convolutional models that have a large temporal receptive field (RF). We argue that these models are not always the best choice for temporal modeling in videos. A large temporal receptive field allows the model to encode the exact sub-action order of a video, which causes a performance decrease when testing videos have a different sub-action order. In this work, we investigate whether we can improve the model robustness to the sub-action order by shrinking the temporal receptive field of action recognition models. For this, we design Video BagNet, a variant of the 3D ResNet-50 model with the temporal receptive field size limited to 1, 9, 17 or 33 frames. We analyze Video BagNet on synthetic and real-world video datasets and experimentally compare models with varying temporal receptive fields. We find that short receptive fields are robust to sub-action order changes, while larger temporal receptive fields are sensitive to the sub-action order.

Jan Warchocki, Teodor Oprescu, Yunhan Wang et al.

In temporal action localization, given an input video, the goal is to predict which actions it contains, where they begin, and where they end. Training and testing current state-of-the-art deep learning models requires access to large amounts of data and computational power. However, gathering such data is challenging and computational resources might be limited. This work explores and measures how current deep temporal action localization models perform in settings constrained by the amount of data or computational power. We measure data efficiency by training each model on a subset of the training set. We find that TemporalMaxer outperforms other models in data-limited settings. Furthermore, we recommend TriDet when training time is limited. To test the efficiency of the models during inference, we pass videos of different lengths through each model. We find that TemporalMaxer requires the least computational resources, likely due to its simple architecture.

Casper van Engelenburg, Seyran Khademi, Jan van Gemert

We propose a simple yet effective metric that measures structural similarity between visual instances of architectural floor plans, without the need for learning. Qualitatively, our experiments show that the retrieval results are similar to deeply learned methods. Effectively comparing instances of floor plan data is paramount to the success of machine understanding of floor plan data, including the assessment of floor plan generative models and floor plan recommendation systems. Comparing visual floor plan images goes beyond a sole pixel-wise visual examination and is crucially about similarities and differences in the shapes and relations between subdivisions that compose the layout. Currently, deep metric learning approaches are used to learn a pair-wise vector representation space that closely mimics the structural similarity, in which the models are trained on similarity labels that are obtained by Intersection-over-Union (IoU). To compensate for the lack of structural awareness in IoU, graph-based approaches such as Graph Matching Networks (GMNs) are used, which require pairwise inference for comparing data instances, making GMNs less practical for retrieval applications. In this paper, an effective evaluation metric for judging the structural similarity of floor plans, coined SSIG (Structural Similarity by IoU and GED), is proposed based on both image and graph distances. In addition, an efficient algorithm is developed that uses SSIG to rank a large-scale floor plan database. Code will be openly available.

Giorgio Cruciata, Luca Cruciata, Liliana Lo Presti et al.

This paper proposes a new method to improve the training efficiency of deep convolutional neural networks. During training, the method evaluates scores to measure how much each layer's parameters change and whether the layer will continue learning or not. Based on these scores, the network is scaled down such that the number of parameters to be learned is reduced, yielding a speed up in training. Unlike state-of-the-art methods that try to compress the network to be used in the inference phase or to limit the number of operations performed in the backpropagation phase, the proposed method is novel in that it focuses on reducing the number of operations performed by the network in the forward propagation during training. The proposed training strategy has been validated on two widely used architecture families: VGG and ResNet. Experiments on MNIST, CIFAR-10 and Imagenette show that, with the proposed method, the training time of the models is more than halved without significantly impacting accuracy. The FLOPs reduction in the forward propagation during training ranges from 17.83\% for VGG-11 to 83.74\% for ResNet-152. These results demonstrate the effectiveness of the proposed technique in speeding up learning of CNNs. The technique will be especially useful in applications where fine-tuning or online training of convolutional models is required, for instance because data arrive sequentially.

Casper van Engelenburg, Fatemeh Mostafavi, Emanuel Kuhn et al.

Diverse and realistic floor plan data are essential for the development of useful computer-aided methods in architectural design. Today's large-scale floor plan datasets predominantly feature simple floor plan layouts, typically representing single-apartment dwellings only. To compensate for the mismatch between current datasets and the real world, we develop \textbf{Modified Swiss Dwellings} (MSD) -- the first large-scale floor plan dataset that contains a significant share of layouts of multi-apartment dwellings. MSD features over 5.3K floor plans of medium- to large-scale building complexes, covering over 18.9K distinct apartments. We validate that existing approaches for floor plan generation, while effective in simpler scenarios, cannot yet seamlessly address the challenges posed by MSD. Our benchmark calls for new research in floor plan machine understanding. Code and data are open.

Attila Lengyel, Ombretta Strafforello, Robert-Jan Bruintjes et al.

Color is a crucial visual cue readily exploited by Convolutional Neural Networks (CNNs) for object recognition. However, CNNs struggle if there is data imbalance between color variations introduced by accidental recording conditions. Color invariance addresses this issue but does so at the cost of removing all color information, which sacrifices discriminative power. In this paper, we propose Color Equivariant Convolutions (CEConvs), a novel deep learning building block that enables shape feature sharing across the color spectrum while retaining important color information. We extend the notion of equivariance from geometric to photometric transformations by incorporating parameter sharing over hue-shifts in a neural network. We demonstrate the benefits of CEConvs in terms of downstream performance to various tasks and improved robustness to color changes, including train-test distribution shifts. Our approach can be seamlessly integrated into existing architectures, such as ResNets, and offers a promising solution for addressing color-based domain shifts in CNNs.

Hesam Araghi, Jan van Gemert, Nergis Tomen

Event cameras offer low-power visual sensing capabilities ideal for edge-device applications. However, their high event rate, driven by high temporal details, can be restrictive in terms of bandwidth and computational resources. In edge AI applications, determining the minimum amount of events for specific tasks can allow reducing the event rate to improve bandwidth, memory, and processing efficiency. In this paper, we study the effect of event subsampling on the accuracy of event data classification using convolutional neural network (CNN) models. Surprisingly, across various datasets, the number of events per video can be reduced by an order of magnitude with little drop in accuracy, revealing the extent to which we can push the boundaries in accuracy vs. event rate trade-off. Additionally, we also find that lower classification accuracy in high subsampling rates is not solely attributable to information loss due to the subsampling of the events, but that the training of CNNs can be challenging in highly subsampled scenarios, where the sensitivity to hyperparameters increases. We quantify training instability across multiple event-based classification datasets using a novel metric for evaluating the hyperparameter sensitivity of CNNs in different subsampling settings. Finally, we analyze the weight gradients of the network to gain insight into this instability.

Ombretta Strafforello, Osman S. Kayhan, Oana Inel et al.

Previous work shows that humans tend to prefer large bounding boxes over small bounding boxes with the same IoU. However, we show here that commonly used object detectors predict large and small boxes equally often. In this work, we investigate how to align automatically detected object boxes with human preference and study whether this improves human quality perception. We evaluate the performance of three commonly used object detectors through a user study (N = 123). We find that humans prefer object detections that are upscaled with factors of 1.5 or 2, even if the corresponding AP is close to 0. Motivated by this result, we propose an asymmetric bounding box regression loss that encourages large over small predicted bounding boxes. Our evaluation study shows that object detectors fine-tuned with the asymmetric loss are better aligned with human preference and are preferred over fixed scaling factors. A qualitative evaluation shows that human preference might be influenced by some object characteristics, like object shape.

Xinqi Li, Yi Zhang, Yidong Zhao et al.

Quantitative cardiac magnetic resonance imaging (MRI) is an increasingly important diagnostic tool for cardiovascular diseases. Yet, co-registration of all baseline images within the quantitative MRI sequence is essential for the accuracy and precision of quantitative maps. However, co-registering all baseline images from a quantitative cardiac MRI sequence remains a nontrivial task because of the simultaneous changes in intensity and contrast, in combination with cardiac and respiratory motion. To address the challenge, we propose a novel motion correction framework based on robust principle component analysis (rPCA) that decomposes quantitative cardiac MRI into low-rank and sparse components, and we integrate the groupwise CNN-based registration backbone within the rPCA framework. The low-rank component of rPCA corresponds to the quantitative mapping (i.e. limited degree of freedom in variation), while the sparse component corresponds to the residual motion, making it easier to formulate and solve the groupwise registration problem. We evaluated our proposed method on cardiac T1 mapping by the modified Look-Locker inversion recovery (MOLLI) sequence, both before and after the Gadolinium contrast agent administration. Our experiments showed that our method effectively improved registration performance over baseline methods without introducing rPCA, and reduced quantitative mapping error in both in-domain (pre-contrast MOLLI) and out-of-domain (post-contrast MOLLI) inference. The proposed rPCA framework is generic and can be integrated with other registration backbones.

Alessandro Duico, Ombretta Strafforello, Jan van Gemert

Predicting which specific parts of a video users will replay is important for several applications, including targeted advertisement placement on video platforms and assisting video creators. In this work, we explore whether it is possible to predict the Most Replayed (MR) data from YouTube videos. To this end, we curate a large video benchmark, the YTMR500 dataset, which comprises 500 YouTube videos with MR data annotations. We evaluate Deep Learning (DL) models of varying complexity on our dataset and perform an extensive ablation study. In addition, we conduct a user study to estimate the human performance on MR data prediction. Our results show that, although by a narrow margin, all the evaluated DL models outperform random predictions. Additionally, they exceed human-level accuracy. This suggests that predicting the MR data is a difficult task that can be enhanced through the assistance of DL. Finally, we believe that DL performance on MR data prediction can be further improved, for example, by using multi-modal learning. We encourage the research community to use our benchmark dataset to further investigate automatic MR data prediction.

Joost Luijmes, Alexander Gielisse, Roman Knyazhitskiy et al.

Implicit neural representations (INRs) encode signals in neural network weights as a memory-efficient representation, decoupling sampling resolution from the associated resource costs. Current INR image classification methods are demonstrated on low-resolution data and are sensitive to image-space transformations. We attribute these issues to the global, fully-connected MLP neural network architecture encoding of current INRs, which lack mechanisms for local representation: MLPs are sensitive to absolute image location and struggle with high-frequency details. We propose ARC: Anchored Representation Clouds, a novel INR architecture that explicitly anchors latent vectors locally in image-space. By introducing spatial structure to the latent vectors, ARC captures local image data which in our testing leads to state-of-the-art implicit image classification of both low- and high-resolution images and increased robustness against image-space translation. Code can be found at https://github.com/JLuij/anchored_representation_clouds.

Hesam Araghi, Jan van Gemert, Nergis Tomen

Event cameras offer high temporal resolution and power efficiency, making them well-suited for edge AI applications. However, their high event rates present challenges for data transmission and processing. Subsampling methods provide a practical solution, but their effect on downstream visual tasks remains underexplored. In this work, we systematically evaluate six hardware-friendly subsampling methods using convolutional neural networks for event video classification on various benchmark datasets. We hypothesize that events from high-density regions carry more task-relevant information and are therefore better suited for subsampling. To test this, we introduce a simple causal density-based subsampling method, demonstrating improved classification accuracy in sparse regimes. Our analysis further highlights key factors affecting subsampling performance, including sensitivity to hyperparameters and failure cases in scenarios with large event count variance. These findings provide insights for utilization of hardware-efficient subsampling strategies that balance data efficiency and task accuracy. The code for this paper will be released at: https://github.com/hesamaraghi/event-camera-subsampling-methods.

Alexander Gielisse, Jan van Gemert

Implicit neural representations (INRs) such as NeRF and SIREN encode a signal in neural network parameters and show excellent results for signal reconstruction. Using INRs for downstream tasks, such as classification, is however not straightforward. Inherent symmetries in the parameters pose challenges and current works primarily focus on designing architectures that are equivariant to these symmetries. However, INR-based classification still significantly under-performs compared to pixel-based methods like CNNs. This work presents an end-to-end strategy for initializing SIRENs together with a learned learning-rate scheme, to yield representations that improve classification accuracy. We show that a simple, straightforward, Transformer model applied to a meta-learned SIREN, without incorporating explicit symmetry equivariances, outperforms the current state-of-the-art. On the CIFAR-10 SIREN classification task, we improve the state-of-the-art without augmentations from 38.8% to 59.6%, and from 63.4% to 64.7% with augmentations. We demonstrate scalability on the high-resolution Imagenette dataset achieving reasonable reconstruction quality with a classification accuracy of 60.8% and are the first to do INR classification on the full ImageNet-1K dataset where we achieve a SIREN classification performance of 23.6%. To the best of our knowledge, no other SIREN classification approach has managed to set a classification baseline for any high-resolution image dataset. Our code is available at https://github.com/SanderGielisse/MWT

Pascal Benschop, Justin Dauwels, Jan van Gemert

Spatial reasoning in vision language models (VLMs) remains fragile when semantics hinge on subtle temporal or geometric cues. We introduce a synthetic benchmark that probes two complementary skills: situational awareness (recognizing whether an interaction is harmful or benign) and spatial awareness (tracking who does what to whom, and reasoning about relative positions and motion). Through minimal video pairs, we test three challenges: distinguishing violence from benign activity, binding assailant roles across viewpoints, and judging fine-grained trajectory alignment. While we evaluate recent VLMs in a training-free setting, the benchmark is applicable to any video classification model. Results show performance only slightly above chance across tasks. A simple aid, stable color cues, partly reduces assailant role confusions but does not resolve the underlying weakness. By releasing data and code, we aim to provide reproducible diagnostics and seed exploration of lightweight spatial priors to complement large-scale pretraining.

Casper van Engelenburg, Jan van Gemert, Seyran Khademi

Floor plans depict building layouts and are often represented as graphs to capture the underlying spatial relationships. Comparison of these graphs is critical for applications like search, clustering, and data visualization. The most successful methods to compare graphs \ie, graph matching networks, rely on costly intermediate cross-graph node-level interactions, therefore being slow in inference time. We introduce \textbf{LayoutGKN}, a more efficient approach that postpones the cross-graph node-level interactions to the end of the joint embedding architecture. We do so by using a differentiable graph kernel as a distance function on the final learned node-level embeddings. We show that LayoutGKN computes similarity comparably or better than graph matching networks while significantly increasing the speed. \href{https://github.com/caspervanengelenburg/LayoutGKN}{Code and data} are open.

Marcos Baptista Rios, Roberto J. López-Sastre, Fabian Caba Heilbron et al.

The Online Action Detection (OAD) problem needs to be revisited. Unlike traditional offline action detection approaches, where the evaluation metrics are clear and well established, in the OAD setting we find very few works and no consensus on the evaluation protocols to be used. In this work we propose to rethink the OAD scenario, clearly defining the problem itself and the main characteristics that the models which are considered online must comply with. We also introduce a novel metric: the Instantaneous Accuracy ($IA$). This new metric exhibits an \emph{online} nature and solves most of the limitations of the previous metrics. We conduct a thorough experimental evaluation on 3 challenging datasets, where the performance of various baseline methods is compared to that of the state-of-the-art. Our results confirm the problems of the previous evaluation protocols, and suggest that an IA-based protocol is more adequate to the online scenario. The baselines models and a development kit with the novel evaluation protocol are publicly available: https://github.com/gramuah/ia.

Marcos Baptista Rios, Roberto J. López-Sastre, Fabian Caba Heilbron et al.

The problem of Online Human Behaviour Recognition in untrimmed videos, aka Online Action Detection (OAD), needs to be revisited. Unlike traditional offline action detection approaches, where the evaluation metrics are clear and well established, in the OAD setting we find few works and no consensus on the evaluation protocols to be used. In this paper we introduce a novel online metric, the Instantaneous Accuracy ($IA$), that exhibits an \emph{online} nature, solving most of the limitations of the previous (offline) metrics. We conduct a thorough experimental evaluation on TVSeries dataset, comparing the performance of various baseline methods to the state of the art. Our results confirm the problems of previous evaluation protocols, and suggest that an IA-based protocol is more adequate to the online scenario for human behaviour understanding. Code of the metric available https://github.com/gramuah/ia

Jan Warchocki, Xi Wang, Jonas Kulhanek et al.

Egocentric video provides a unique view into human perception and interaction, with growing relevance for augmented reality, robotics, and assistive technologies. However, rapid camera motion and complex scene dynamics pose major challenges for 3D reconstruction from this perspective. While 3D Gaussian Splatting (3DGS) has become a state-of-the-art method for efficient, high-quality novel view synthesis, variants, that focus on reconstructing dynamic scenes from monocular video are rarely evaluated on egocentric video. It remains unclear whether existing models generalize to this setting or if egocentric-specific solutions are needed. In this work, we evaluate dynamic monocular 3DGS models on egocentric and exocentric video using paired ego-exo recordings from the EgoExo4D dataset. We find that reconstruction quality is consistently lower in egocentric views. Analysis reveals that the difference in reconstruction quality, measured in peak signal-to-noise ratio, stems from the reconstruction of static, not dynamic, content. Our findings underscore current limitations and motivate the development of egocentric-specific approaches, while also highlighting the value of separately evaluating static and dynamic regions of a video.

Zhi-Yi Lin, Jouh Yeong Chew, Jan van Gemert et al.

Precisely detecting which object a person is paying attention to is critical for human-robot interaction since it provides important cues for the next action from the human user. We propose an end-to-end approach for gaze target detection: predicting a head-target connection between individuals and the target image regions they are looking at. Most of the existing methods use independent components such as off-the-shelf head detectors or have problems in establishing associations between heads and gaze targets. In contrast, we investigate an end-to-end multi-person Gaze target detection framework with Heads and Targets Association (GazeHTA), which predicts multiple head-target instances based solely on input scene image. GazeHTA addresses challenges in gaze target detection by (1) leveraging a pre-trained diffusion model to extract scene features for rich semantic understanding, (2) re-injecting a head feature to enhance the head priors for improved head understanding, and (3) learning a connection map as the explicit visual associations between heads and gaze targets. Our extensive experimental results demonstrate that GazeHTA outperforms state-of-the-art gaze target detection methods and two adapted diffusion-based baselines on two standard datasets.

Alejandro Castañeda Garcia, Jan van Gemert, Daan Brinks et al.

Autoencoders can be challenged by spatially non-uniform sampling of image content. This is common in medical imaging, biology, and physics, where informative patterns occur rarely at specific image coordinates, as background dominates these locations in most samples, biasing reconstructions toward the majority appearance. In practice, autoencoders are biased toward dominant patterns resulting in the loss of fine-grained detail and causing blurred reconstructions for rare spatial inputs especially under spatial data imbalance. We address spatial imbalance by two complementary components: (i) self-entropy-based loss that upweights statistically uncommon spatial locations and (ii) Sample Propagation, a replay mechanism that selectively re-exposes the model to hard to reconstruct samples across batches during training. We benchmark existing data balancing strategies, originally developed for supervised classification, in the unsupervised reconstruction setting. Drawing on the limitations of these approaches, our method specifically targets spatial imbalance by encouraging models to focus on statistically rare locations, improving reconstruction consistency compared to existing baselines. We validate in a simulated dataset with controlled spatial imbalance conditions, and in three, uncontrolled, diverse real-world datasets spanning physical, biological, and astronomical domains. Our approach outperforms baselines on various reconstruction metrics, particularly under spatial imbalance distributions. These results highlight the importance of data representation in a batch and emphasize rare samples in unsupervised image reconstruction. We will make all code and related data available.

Robert-Jan Bruintjes, Attila Lengyel, Osman Semih Kayhan et al.

Deep Learning requires large amounts of data to train models that work well. In data-deficient settings, performance can be degraded. We investigate which Deep Learning methods benefit training models in a data-deficient setting, by organizing the "VIPriors: Visual Inductive Priors for Data-Efficient Deep Learning" workshop series, featuring four editions of data-impaired challenges. These challenges address the problem of training deep learning models for computer vision tasks with limited data. Participants are limited to training models from scratch using a low number of training samples and are not allowed to use any form of transfer learning. We aim to stimulate the development of novel approaches that incorporate prior knowledge to improve the data efficiency of deep learning models. Successful challenge entries make use of large model ensembles that mix Transformers and CNNs, as well as heavy data augmentation. Novel prior knowledge-based methods contribute to success in some entries.

Robert-Jan Bruintjes, Jan van Gemert

How discriminative position information is for image classification depends on the data. On the one hand, the camera position is arbitrary and objects can appear anywhere in the image, arguing for translation invariance. At the same time, position information is key for exploiting capture/center bias, and scene layout, e.g.: the sky is up. We show that position bias, the level to which a dataset is more easily solved when positional information on input features is used, plays a crucial role in the performance of Vision Transformers image classifiers. To investigate, we propose Position-SHAP, a direct measure of position bias by extending SHAP to work with position embeddings. We show various levels of position bias in different datasets, and find that the optimal choice of position embedding depends on the position bias apparent in the dataset. We therefore propose Auto-PE, a single-parameter position embedding extension, which allows the position embedding to modulate its norm, enabling the unlearning of position information. Auto-PE combines with existing PEs to match or improve accuracy on classification datasets.

Alexander Gielisse, Nergis Tömen, Jan van Gemert

Most recent works on optical flow use convex upsampling as the last step to obtain high-resolution flow. In this work, we show and discuss several issues and limitations of this currently widely adopted convex upsampling approach. We propose a series of changes, in an attempt to resolve current issues. First, we propose to decouple the weights for the final convex upsampler, making it easier to find the correct convex combination. For the same reason, we also provide extra contextual features to the convex upsampler. Then, we increase the convex mask size by using an attention-based alternative convex upsampler; Transformers for Convex Upsampling. This upsampler is based on the observation that convex upsampling can be reformulated as attention, and we propose to use local attention masks as a drop-in replacement for convex masks to increase the mask size. We provide empirical evidence that a larger mask size increases the likelihood of the existence of the convex combination. Lastly, we propose an alternative training scheme to remove bilinear interpolation artifacts from the model output. Our proposed ideas could theoretically be applied to almost every current state-of-the-art optical flow architecture. On the FlyingChairs + FlyingThings3D training setting we reduce the Sintel Clean training end-point-error of RAFT from 1.42 to 1.26, GMA from 1.31 to 1.18, and that of FlowFormer from 0.94 to 0.90, by solely adapting the convex upsampler.

Robert-Jan Bruintjes, Attila Lengyel, Marcos Baptista Rios et al.

The fourth edition of the "VIPriors: Visual Inductive Priors for Data-Efficient Deep Learning" workshop features two data-impaired challenges. These challenges address the problem of training deep learning models for computer vision tasks with limited data. Participants are limited to training models from scratch using a low number of training samples and are not allowed to use any form of transfer learning. We aim to stimulate the development of novel approaches that incorporate inductive biases to improve the data efficiency of deep learning models. Significant advancements are made compared to the provided baselines, where winning solutions surpass the baselines by a considerable margin in both tasks. As in previous editions, these achievements are primarily attributed to heavy use of data augmentation policies and large model ensembles, though novel prior-based methods seem to contribute more to successful solutions compared to last year. This report highlights the key aspects of the challenges and their outcomes.

Sven de Witte, Ombretta Strafforello, Jan van Gemert

Bounding boxes are often used to communicate automatic object detection results to humans, aiding humans in a multitude of tasks. We investigate the relationship between bounding box localization errors and human task performance. We use observer performance studies on a visual multi-object counting task to measure both human trust and performance with different levels of bounding box accuracy. The results show that localization errors have no significant impact on human accuracy or trust in the system. Recall and precision errors impact both human performance and trust, suggesting that optimizing algorithms based on the F1 score is more beneficial in human-computer tasks. Lastly, the paper offers an improvement on bounding boxes in multi-object counting tasks with center dots, showing improved performance and better resilience to localization inaccuracy.

Robert-Jan Bruintjes, Attila Lengyel, Marcos Baptista Rios et al.

The third edition of the "VIPriors: Visual Inductive Priors for Data-Efficient Deep Learning" workshop featured four data-impaired challenges, focusing on addressing the limitations of data availability in training deep learning models for computer vision tasks. The challenges comprised of four distinct data-impaired tasks, where participants were required to train models from scratch using a reduced number of training samples. The primary objective was to encourage novel approaches that incorporate relevant inductive biases to enhance the data efficiency of deep learning models. To foster creativity and exploration, participants were strictly prohibited from utilizing pre-trained checkpoints and other transfer learning techniques. Significant advancements were made compared to the provided baselines, where winning solutions surpassed the baselines by a considerable margin in all four tasks. These achievements were primarily attributed to the effective utilization of extensive data augmentation policies, model ensembling techniques, and the implementation of data-efficient training methods, including self-supervised representation learning. This report highlights the key aspects of the challenges and their outcomes.

Burak Yildiz, Seyran Khademi, Ronald Maria Siebes et al.

We introduce AmsterTime: a challenging dataset to benchmark visual place recognition (VPR) in presence of a severe domain shift. AmsterTime offers a collection of 2,500 well-curated images matching the same scene from a street view matched to historical archival image data from Amsterdam city. The image pairs capture the same place with different cameras, viewpoints, and appearances. Unlike existing benchmark datasets, AmsterTime is directly crowdsourced in a GIS navigation platform (Mapillary). We evaluate various baselines, including non-learning, supervised and self-supervised methods, pre-trained on different relevant datasets, for both verification and retrieval tasks. Our result credits the best accuracy to the ResNet-101 model pre-trained on the Landmarks dataset for both verification and retrieval tasks by 84% and 24%, respectively. Additionally, a subset of Amsterdam landmarks is collected for feature evaluation in a classification task. Classification labels are further used to extract the visual explanations using Grad-CAM for inspection of the learned similar visuals in a deep metric learning models.

Attila Lengyel, Robert-Jan Bruintjes, Marcos Baptista Rios et al.

The second edition of the "VIPriors: Visual Inductive Priors for Data-Efficient Deep Learning" challenges featured five data-impaired challenges, where models are trained from scratch on a reduced number of training samples for various key computer vision tasks. To encourage new and creative ideas on incorporating relevant inductive biases to improve the data efficiency of deep learning models, we prohibited the use of pre-trained checkpoints and other transfer learning techniques. The provided baselines are outperformed by a large margin in all five challenges, mainly thanks to extensive data augmentation policies, model ensembling, and data efficient network architectures.

Yancong Lin, Silvia-Laura Pintea, Jan van Gemert

Many objects are naturally symmetric, and this symmetry can be exploited to infer unseen 3D properties from a single 2D image. Recently, NeRD is proposed for accurate 3D mirror plane estimation from a single image. Despite the unprecedented accuracy, it relies on large annotated datasets for training and suffers from slow inference. Here we aim to improve its data and compute efficiency. We do away with the computationally expensive 4D feature volumes and instead explicitly compute the feature correlation of the pixel correspondences across depth, thus creating a compact 3D volume. We also design multi-stage spherical convolutions to identify the optimal mirror plane on the hemisphere, whose inductive bias offers gains in data-efficiency. Experiments on both synthetic and real-world datasets show the benefit of our proposed changes for improved data efficiency and inference speed.

Tuhin Das, Robert-Jan Bruintjes, Attila Lengyel et al.

To alleviate lower classification performance on rare classes in imbalanced datasets, a possible solution is to augment the underrepresented classes with synthetic samples. Domain adaptation can be incorporated in a classifier to decrease the domain discrepancy between real and synthetic samples. While domain adaptation is generally applied on completely synthetic source domains and real target domains, we explore how domain adaptation can be applied when only a single rare class is augmented with simulated samples. As a testbed, we use a camera trap animal dataset with a rare deer class, which is augmented with synthetic deer samples. We adapt existing domain adaptation methods to two new methods for the single rare class setting: DeerDANN, based on the Domain-Adversarial Neural Network (DANN), and DeerCORAL, based on deep correlation alignment (Deep CORAL) architectures. Experiments show that DeerDANN has the highest improvement in deer classification accuracy of 24.0% versus 22.4% improvement of DeerCORAL when compared to the baseline. Further, both methods require fewer than 10k synthetic samples, as used by the baseline, to achieve these higher accuracies. DeerCORAL requires the least number of synthetic samples (2k deer), followed by DeerDANN (8k deer).

Yancong Lin, Silvia-Laura Pintea, Jan van Gemert

Current work on lane detection relies on large manually annotated datasets. We reduce the dependency on annotations by leveraging massive cheaply available unlabelled data. We propose a novel loss function exploiting geometric knowledge of lanes in Hough space, where a lane can be identified as a local maximum. By splitting lanes into separate channels, we can localize each lane via simple global max-pooling. The location of the maximum encodes the layout of a lane, while the intensity indicates the the probability of a lane being present. Maximizing the log-probability of the maximal bins helps neural networks find lanes without labels. On the CULane and TuSimple datasets, we show that the proposed Hough Transform loss improves performance significantly by learning from large amounts of unlabelled images.

Pranjal Singh Rajput, Yeshwanth Napolean, Jan van Gemert

Annotating a large-scale in-the-wild person re-identification dataset especially of marathon runners is a challenging task. The variations in the scenarios such as camera viewpoints, resolution, occlusion, and illumination make the problem non-trivial. Manually annotating bounding boxes in such large-scale datasets is cost-inefficient. Additionally, due to crowdedness and occlusion in the videos, aligning the identity of runners across multiple disjoint cameras is a challenge. We collected a novel large-scale in-the-wild video dataset of marathon runners. The dataset consists of hours of recording of thousands of runners captured using 42 hand-held smartphone cameras and covering real-world scenarios. Due to the presence of crowdedness and occlusion in the videos, the annotation of runners becomes a challenging task. We propose a new scheme for tackling the challenges in the annotation of such large dataset. Our technique reduces the overall cost of annotation in terms of time as well as budget. We demonstrate performing fps analysis to reduce the effort and time of annotation. We investigate several annotation methods for efficiently generating tight bounding boxes. Our results prove that interpolating bounding boxes between keyframes is the most efficient method of bounding box generation amongst several other methods and is 3x times faster than the naive baseline method. We introduce a novel way of aligning the identity of runners in disjoint cameras. Our inter-camera alignment tool integrated with the state-of-the-art person re-id system proves to be sufficient and effective in the alignment of the runners across multiple cameras with non-overlapping views. Our proposed framework of annotation reduces the annotation cost of the dataset by a factor of 16x, also effectively aligning 93.64% of the runners in the cross-camera setting.

Robert-Jan Bruintjes, Attila Lengyel, Marcos Baptista Rios et al.

We present the first edition of "VIPriors: Visual Inductive Priors for Data-Efficient Deep Learning" challenges. We offer four data-impaired challenges, where models are trained from scratch, and we reduce the number of training samples to a fraction of the full set. Furthermore, to encourage data efficient solutions, we prohibited the use of pre-trained models and other transfer learning techniques. The majority of top ranking solutions make heavy use of data augmentation, model ensembling, and novel and efficient network architectures to achieve significant performance increases compared to the provided baselines.

Nergis Tomen, Jan van Gemert

Convolutional layers in CNNs implement linear filters which decompose the input into different frequency bands. However, most modern architectures neglect standard principles of filter design when optimizing their model choices regarding the size and shape of the convolutional kernel. In this work, we consider the well-known problem of spectral leakage caused by windowing artifacts in filtering operations in the context of CNNs. We show that the small size of CNN kernels make them susceptible to spectral leakage, which may induce performance-degrading artifacts. To address this issue, we propose the use of larger kernel sizes along with the Hamming window function to alleviate leakage in CNN architectures. We demonstrate improved classification accuracy on multiple benchmark datasets including Fashion-MNIST, CIFAR-10, CIFAR-100 and ImageNet with the simple use of a standard window function in convolutional layers. Finally, we show that CNNs employing the Hamming window display increased robustness against various adversarial attacks.

Amogh Gudi, Marian Bittner, Jan van Gemert

Remote photo-plethysmography (rPPG) uses a camera to estimate a person's heart rate (HR). Similar to how heart rate can provide useful information about a person's vital signs, insights about the underlying physio/psychological conditions can be obtained from heart rate variability (HRV). HRV is a measure of the fine fluctuations in the intervals between heart beats. However, this measure requires temporally locating heart beats with a high degree of precision. We introduce a refined and efficient real-time rPPG pipeline with novel filtering and motion suppression that not only estimates heart rates, but also extracts the pulse waveform to time heart beats and measure heart rate variability. This unsupervised method requires no rPPG specific training and is able to operate in real-time. We also introduce a new multi-modal video dataset, VicarPPG 2, specifically designed to evaluate rPPG algorithms on HR and HRV estimation. We validate and study our method under various conditions on a comprehensive range of public and self-recorded datasets, showing state-of-the-art results and providing useful insights into some unique aspects. Lastly, we make available CleanerPPG, a collection of human-verified ground truth peak/heart-beat annotations for existing rPPG datasets. These verified annotations should make future evaluations and benchmarking of rPPG algorithms more accurate, standardized and fair.

Yunqiang Li, Jan van Gemert

Unsupervised hashing is important for indexing huge image or video collections without having expensive annotations available. Hashing aims to learn short binary codes for compact storage and efficient semantic retrieval. We propose an unsupervised deep hashing layer called Bi-half Net that maximizes entropy of the binary codes. Entropy is maximal when both possible values of the bit are uniformly (half-half) distributed. To maximize bit entropy, we do not add a term to the loss function as this is difficult to optimize and tune. Instead, we design a new parameter-free network layer to explicitly force continuous image features to approximate the optimal half-half bit distribution. This layer is shown to minimize a penalized term of the Wasserstein distance between the learned continuous image features and the optimal half-half bit distribution. Experimental results on the image datasets Flickr25k, Nus-wide, Cifar-10, Mscoco, Mnist and the video datasets Ucf-101 and Hmdb-51 show that our approach leads to compact codes and compares favorably to the current state-of-the-art.