Sinan Kalkan

Fehmi Kahraman, Kemal Oksuz, Sinan Kalkan et al.

Object detectors are conventionally trained by a weighted sum of classification and localization losses. Recent studies (e.g., predicting IoU with an auxiliary head, Generalized Focal Loss, Rank & Sort Loss) have shown that forcing these two loss terms to interact with each other in non-conventional ways creates a useful inductive bias and improves performance. Inspired by these works, we focus on the correlation between classification and localization and make two main contributions: (i) We provide an analysis about the effects of correlation between classification and localization tasks in object detectors. We identify why correlation affects the performance of various NMS-based and NMS-free detectors, and we devise measures to evaluate the effect of correlation and use them to analyze common detectors. (ii) Motivated by our observations, e.g., that NMS-free detectors can also benefit from correlation, we propose Correlation Loss, a novel plug-in loss function that improves the performance of various object detectors by directly optimizing correlation coefficients: E.g., Correlation Loss on Sparse R-CNN, an NMS-free method, yields 1.6 AP gain on COCO and 1.8 AP gain on Cityscapes dataset. Our best model on Sparse R-CNN reaches 51.0 AP without test-time augmentation on COCO test-dev, reaching state-of-the-art. Code is available at https://github.com/fehmikahraman/CorrLoss

Özgür Aslan, Burak Bolat, Batuhan Bal et al.

The rise of simulation environments has enabled learning-based approaches for assembly planning, which is otherwise a labor-intensive and daunting task. Assembling furniture is especially interesting since furniture are intricate and pose challenges for learning-based approaches. Surprisingly, humans can solve furniture assembly mostly given a 2D snapshot of the assembled product. Although recent years have witnessed promising learning-based approaches for furniture assembly, they assume the availability of correct connection labels for each assembly step, which are expensive to obtain in practice. In this paper, we alleviate this assumption and aim to solve furniture assembly with as little human expertise and supervision as possible. To be specific, we assume the availability of the assembled point cloud, and comparing the point cloud of the current assembly and the point cloud of the target product, obtain a novel reward signal based on two measures: Incorrectness and incompleteness. We show that our novel reward signal can train a deep network to successfully assemble different types of furniture. Code and networks available here: https://github.com/METU-KALFA/AssembleRL

Feyza Yavuz, Baris Can Cam, Adnan Harun Dogan et al.

Ranking-based loss functions, such as Average Precision Loss and Rank&Sort Loss, outperform widely used score-based losses in object detection. These loss functions better align with the evaluation criteria, have fewer hyperparameters, and offer robustness against the imbalance between positive and negative classes. However, they require pairwise comparisons among $P$ positive and $N$ negative predictions, introducing a time complexity of $\mathcal{O}(PN)$, which is prohibitive since $N$ is often large (e.g., $10^8$ in ATSS). Despite their advantages, the widespread adoption of ranking-based losses has been hindered by their high time and space complexities. In this paper, we focus on improving the efficiency of ranking-based loss functions. To this end, we propose Bucketed Ranking-based Losses which group negative predictions into $B$ buckets ($B \ll N$) in order to reduce the number of pairwise comparisons so that time complexity can be reduced. Our method enhances the time complexity, reducing it to $\mathcal{O}(\max (N \log(N), P^2))$. To validate our method and show its generality, we conducted experiments on 2 different tasks, 3 different datasets, 7 different detectors. We show that Bucketed Ranking-based (BR) Losses yield the same accuracy with the unbucketed versions and provide $2\times$ faster training on average. We also train, for the first time, transformer-based object detectors using ranking-based losses, thanks to the efficiency of our BR. When we train CoDETR, a state-of-the-art transformer-based object detector, using our BR Loss, we consistently outperform its original results over several different backbones. Code is available at https://github.com/blisgard/BucketedRankingBasedLosses

Çağrı Eser, Zeynep Sonat Baltacı, Emre Akbaş et al.

Imbalance in classification tasks is commonly quantified by the cardinalities of examples across classes. This, however, disregards the presence of redundant examples and inherent differences in the learning difficulties of classes. Alternatively, one can use complex measures such as training loss and uncertainty, which, however, depend on training a machine learning model. Our paper proposes using data Intrinsic Dimensionality (ID) as an easy-to-compute, model-free measure of imbalance that can be seamlessly incorporated into various imbalance mitigation methods. Our results across five different datasets with a diverse range of imbalance ratios show that ID consistently outperforms cardinality-based re-weighting and re-sampling techniques used in the literature. Moreover, we show that combining ID with cardinality can further improve performance. Code: https://github.com/cagries/IDIM.

Burak Akgül, Erol Şahin, Sinan Kalkan

While appearance-based gaze estimation has achieved significant improvements in accuracy and domain adaptation, the fairness of these systems across different demographic groups remains largely unexplored. To date, there is no comprehensive benchmark quantifying algorithmic bias in gaze estimation. This paper presents the first extensive evaluation of fairness in appearance-based gaze estimation, focusing on ethnicity and gender attributes. We establish a fairness baseline by analyzing state-of-the-art models using standard fairness metrics, revealing significant performance disparities. Furthermore, we evaluate the effectiveness of existing bias mitigation strategies when applied to the gaze domain and show that their fairness contributions are limited. We summarize key insights and open issues. Overall, our work calls for research into developing robust, equitable gaze estimators. To support future research and reproducibility, we publicly release our annotations, code, and trained models at: github.com/akgulburak/gaze-estimation-fairness

Bedrettin Cetinkaya, Sinan Kalkan, Emre Akbas

Ground-truth depth, when combined with color data, helps improve object detection accuracy over baseline models that only use color. However, estimated depth does not always yield improvements. Many factors affect the performance of object detection when estimated depth is used. In this paper, we comprehensively investigate these factors with detailed experiments, such as using ground-truth vs. estimated depth, effects of different state-of-the-art depth estimation networks, effects of using different indoor and outdoor RGB-D datasets as training data for depth estimation, and different architectural choices for integrating depth to the base object detector network. We propose an early concatenation strategy of depth, which yields higher mAP than previous works' while using significantly fewer parameters.

Bedrettin Cetinkaya, Sinan Kalkan, Emre Akbas

Generating crisp, i.e., one-pixel-wide, edge maps remains one of the fundamental challenges in edge detection, affecting both traditional and learning-based methods. To obtain crisp edges, most existing approaches rely on two hand-crafted post-processing algorithms, Non-Maximum Suppression (NMS) and skeleton-based thinning, which are non-differentiable and hinder end-to-end optimization. Moreover, all existing crisp edge detection methods still depend on such post-processing to achieve satisfactory results. To address this limitation, we propose \MethodLPP, a lightweight, only $\sim$21K additional parameters, and plug-and-play matching-based supervision module that can be appended to any edge detection model for joint end-to-end learning of crisp edges. At each training iteration, \MethodLPP performs one-to-one matching between predicted and ground-truth edges based on spatial distance and confidence, ensuring consistency between training and testing protocols. Extensive experiments on four popular datasets demonstrate that integrating \MethodLPP substantially improves the performance of existing edge detection models. In particular, \MethodLPP increases the Average Crispness (AC) metric by up to 2--4$\times$ compared to baseline models. Under the crispness-emphasized evaluation (CEval), \MethodLPP further boosts baseline performance by up to 20--35\% in ODS and achieves similar gains in OIS and AP, achieving SOTA performance that matches or surpasses standard post-processing for the first time. Code is available at https://cvpr26-matched.github.io.

Feyza Yavuz, Sinan Kalkan

Logo retrieval is a challenging problem since the definition of similarity is more subjective compared to image retrieval tasks and the set of known similarities is very scarce. To tackle this challenge, in this paper, we propose a simple but effective segment-based augmentation strategy to introduce artificially similar logos for training deep networks for logo retrieval. In this novel augmentation strategy, we first find segments in a logo and apply transformations such as rotation, scaling, and color change, on the segments, unlike the conventional image-level augmentation strategies. Moreover, we evaluate whether the recently introduced ranking-based loss function, Smooth-AP, is a better approach for learning similarity for logo retrieval. On the large scale METU Trademark Dataset, we show that (i) our segment-based augmentation strategy improves retrieval performance compared to the baseline model or image-level augmentation strategies, and (ii) Smooth-AP indeed performs better than conventional losses for logo retrieval.

Bedrettin Cetinkaya, Sinan Kalkan, Emre Akbas

Detecting edges in images suffers from the problems of (P1) heavy imbalance between positive and negative classes as well as (P2) label uncertainty owing to disagreement between different annotators. Existing solutions address P1 using class-balanced cross-entropy loss and dice loss and P2 by only predicting edges agreed upon by most annotators. In this paper, we propose RankED, a unified ranking-based approach that addresses both the imbalance problem (P1) and the uncertainty problem (P2). RankED tackles these two problems with two components: One component which ranks positive pixels over negative pixels, and the second which promotes high confidence edge pixels to have more label certainty. We show that RankED outperforms previous studies and sets a new state-of-the-art on NYUD-v2, BSDS500 and Multi-cue datasets. Code is available at https://ranked-cvpr24.github.io.

Berk Sezer, Ali Görkem Küçük, Erol Şahin et al.

While zero-shot appearance-based 3D gaze estimation offers significant cost-efficiency by directly mapping RGB images to gaze vectors, its reliability in Human-Robot Interaction (HRI) settings remains uncertain. Existing benchmarks frequently overlook fundamental HRI conditions, such as dynamic camera viewpoints and moving targets in video. Furthermore, current cross-dataset evaluations often suffer from a complexity gap, where methods trained on diverse datasets are tested on significantly smaller and less varied sets, failing to assess true robustness. To bridge these gaps, we introduce Gaze4HRI, a large-scale dataset (50+ subjects, 3,000+ videos, 600,000+ frames) designed to evaluate state-of-the-art performance against critical HRI variables: illumination, head-gaze conflict, as well as the motion of camera and gaze target in video. Our benchmark reveals that all evaluated methods fail in at least one condition, identifying steeply-downward gaze as a universal failure point. Notably, PureGaze trained on the ETH-X-Gaze dataset uniquely maintains resilience across all other conditions. These results challenge the recent focus in the literature on complex spatial-temporal modeling and Transformer-based architectures. Instead, our findings suggest that extensive data diversity, as exemplified by the ETH-X-Gaze dataset, serves as the primary driver of zero-shot robustness in unconstrained environments, while resilience-enhancing frameworks, such as PureGaze's self-adversarial loss for gaze feature purification, provide a substantial further improvement. Ultimately, this study establishes a rigorous benchmark that provides practical guidelines for practitioners as well as reshaping future research. The dataset and codes are available at https://gazeforhri.github.io.

Kemal Oksuz, Baris Can Cam, Fehmi Kahraman et al.

This paper presents Mask-aware Intersection-over-Union (maIoU) for assigning anchor boxes as positives and negatives during training of instance segmentation methods. Unlike conventional IoU or its variants, which only considers the proximity of two boxes; maIoU consistently measures the proximity of an anchor box with not only a ground truth box but also its associated ground truth mask. Thus, additionally considering the mask, which, in fact, represents the shape of the object, maIoU enables a more accurate supervision during training. We present the effectiveness of maIoU on a state-of-the-art (SOTA) assigner, ATSS, by replacing IoU operation by our maIoU and training YOLACT, a SOTA real-time instance segmentation method. Using ATSS with maIoU consistently outperforms (i) ATSS with IoU by $\sim 1$ mask AP, (ii) baseline YOLACT with fixed IoU threshold assigner by $\sim 2$ mask AP over different image sizes and (iii) decreases the inference time by $25 \%$ owing to using less anchors. Then, exploiting this efficiency, we devise maYOLACT, a faster and $+6$ AP more accurate detector than YOLACT. Our best model achieves $37.7$ mask AP at $25$ fps on COCO test-dev establishing a new state-of-the-art for real-time instance segmentation. Code is available at https://github.com/kemaloksuz/Mask-aware-IoU

Kemal Oksuz, Baris Can Cam, Emre Akbas et al.

We propose Rank & Sort (RS) Loss, a ranking-based loss function to train deep object detection and instance segmentation methods (i.e. visual detectors). RS Loss supervises the classifier, a sub-network of these methods, to rank each positive above all negatives as well as to sort positives among themselves with respect to (wrt.) their localisation qualities (e.g. Intersection-over-Union - IoU). To tackle the non-differentiable nature of ranking and sorting, we reformulate the incorporation of error-driven update with backpropagation as Identity Update, which enables us to model our novel sorting error among positives. With RS Loss, we significantly simplify training: (i) Thanks to our sorting objective, the positives are prioritized by the classifier without an additional auxiliary head (e.g. for centerness, IoU, mask-IoU), (ii) due to its ranking-based nature, RS Loss is robust to class imbalance, and thus, no sampling heuristic is required, and (iii) we address the multi-task nature of visual detectors using tuning-free task-balancing coefficients. Using RS Loss, we train seven diverse visual detectors only by tuning the learning rate, and show that it consistently outperforms baselines: e.g. our RS Loss improves (i) Faster R-CNN by ~ 3 box AP and aLRP Loss (ranking-based baseline) by ~ 2 box AP on COCO dataset, (ii) Mask R-CNN with repeat factor sampling (RFS) by 3.5 mask AP (~ 7 AP for rare classes) on LVIS dataset; and also outperforms all counterparts. Code is available at: https://github.com/kemaloksuz/RankSortLoss

Kemal Oksuz, Baris Can Cam, Sinan Kalkan et al.

Despite being widely used as a performance measure for visual detection tasks, Average Precision (AP) is limited in (i) reflecting localisation quality, (ii) interpretability and (iii) robustness to the design choices regarding its computation, and its applicability to outputs without confidence scores. Panoptic Quality (PQ), a measure proposed for evaluating panoptic segmentation (Kirillov et al., 2019), does not suffer from these limitations but is limited to panoptic segmentation. In this paper, we propose Localisation Recall Precision (LRP) Error as the average matching error of a visual detector computed based on both its localisation and classification qualities for a given confidence score threshold. LRP Error, initially proposed only for object detection by Oksuz et al. (2018), does not suffer from the aforementioned limitations and is applicable to all visual detection tasks. We also introduce Optimal LRP (oLRP) Error as the minimum LRP Error obtained over confidence scores to evaluate visual detectors and obtain optimal thresholds for deployment. We provide a detailed comparative analysis of LRP Error with AP and PQ, and use nearly 100 state-of-the-art visual detectors from seven visual detection tasks (i.e. object detection, keypoint detection, instance segmentation, panoptic segmentation, visual relationship detection, zero-shot detection and generalised zero-shot detection) using ten datasets to empirically show that LRP Error provides richer and more discriminative information than its counterparts. Code available at: https://github.com/kemaloksuz/LRP-Error

Kemal Oksuz, Baris Can Cam, Emre Akbas et al.

We propose average Localisation-Recall-Precision (aLRP), a unified, bounded, balanced and ranking-based loss function for both classification and localisation tasks in object detection. aLRP extends the Localisation-Recall-Precision (LRP) performance metric (Oksuz et al., 2018) inspired from how Average Precision (AP) Loss extends precision to a ranking-based loss function for classification (Chen et al., 2020). aLRP has the following distinct advantages: (i) aLRP is the first ranking-based loss function for both classification and localisation tasks. (ii) Thanks to using ranking for both tasks, aLRP naturally enforces high-quality localisation for high-precision classification. (iii) aLRP provides provable balance between positives and negatives. (iv) Compared to on average $\sim$6 hyperparameters in the loss functions of state-of-the-art detectors, aLRP Loss has only one hyperparameter, which we did not tune in practice. On the COCO dataset, aLRP Loss improves its ranking-based predecessor, AP Loss, up to around $5$ AP points, achieves $48.9$ AP without test time augmentation and outperforms all one-stage detectors. Code available at: https://github.com/kemaloksuz/aLRPLoss .

Kemal Oksuz, Baris Can Cam, Emre Akbas et al.

Two-stage deep object detectors generate a set of regions-of-interest (RoI) in the first stage, then, in the second stage, identify objects among the proposed RoIs that sufficiently overlap with a ground truth (GT) box. The second stage is known to suffer from a bias towards RoIs that have low intersection-over-union (IoU) with the associated GT boxes. To address this issue, we first propose a sampling method to generate bounding boxes (BB) that overlap with a given reference box more than a given IoU threshold. Then, we use this BB generation method to develop a positive RoI (pRoI) generator that produces RoIs following any desired spatial or IoU distribution, for the second-stage. We show that our pRoI generator is able to simulate other sampling methods for positive examples such as hard example mining and prime sampling. Using our generator as an analysis tool, we show that (i) IoU imbalance has an adverse effect on performance, (ii) hard positive example mining improves the performance only for certain input IoU distributions, and (iii) the imbalance among the foreground classes has an adverse effect on performance and that it can be alleviated at the batch level. Finally, we train Faster R-CNN using our pRoI generator and, compared to conventional training, obtain better or on-par performance for low IoUs and significant improvements when trained for higher IoUs for Pascal VOC and MS COCO datasets. The code is available at: https://github.com/kemaloksuz/BoundingBoxGenerator.

Kemal Oksuz, Baris Can Cam, Sinan Kalkan et al.

In this paper, we present a comprehensive review of the imbalance problems in object detection. To analyze the problems in a systematic manner, we introduce a problem-based taxonomy. Following this taxonomy, we discuss each problem in depth and present a unifying yet critical perspective on the solutions in the literature. In addition, we identify major open issues regarding the existing imbalance problems as well as imbalance problems that have not been discussed before. Moreover, in order to keep our review up to date, we provide an accompanying webpage which catalogs papers addressing imbalance problems, according to our problem-based taxonomy. Researchers can track newer studies on this webpage available at: https://github.com/kemaloksuz/ObjectDetectionImbalance .

Hazan Anayurt, Sezai Artun Ozyegin, Ulfet Cetin et al.

Humans frequently use referring (identifying) expressions to refer to objects. Especially in ambiguous settings, humans prefer expressions (called relational referring expressions) that describe an object with respect to a distinguishing, unique object. Unlike studies on video object search using referring expressions, in this paper, our focus is on (i) relational referring expressions in highly ambiguous settings, and (ii) methods that can both generate and comprehend a referring expression. For this goal, we first introduce a new dataset for video object search with referring expressions that includes numerous copies of the objects, making it difficult to use non-relational expressions. Moreover, we train two baseline deep networks on this dataset, which show promising results. Finally, we propose a deep attention network that significantly outperforms the baselines on our dataset. The dataset and the codes are available at https://github.com/hazananayurt/viref.

Kemal Oksuz, Baris Can Cam, Emre Akbas et al.

Average precision (AP), the area under the recall-precision (RP) curve, is the standard performance measure for object detection. Despite its wide acceptance, it has a number of shortcomings, the most important of which are (i) the inability to distinguish very different RP curves, and (ii) the lack of directly measuring bounding box localization accuracy. In this paper, we propose 'Localization Recall Precision (LRP) Error', a new metric which we specifically designed for object detection. LRP Error is composed of three components related to localization, false negative (FN) rate and false positive (FP) rate. Based on LRP, we introduce the 'Optimal LRP', the minimum achievable LRP error representing the best achievable configuration of the detector in terms of recall-precision and the tightness of the boxes. In contrast to AP, which considers precisions over the entire recall domain, Optimal LRP determines the 'best' confidence score threshold for a class, which balances the trade-off between localization and recall-precision. In our experiments, we show that, for state-of-the-art object (SOTA) detectors, Optimal LRP provides richer and more discriminative information than AP. We also demonstrate that the best confidence score thresholds vary significantly among classes and detectors. Moreover, we present LRP results of a simple online video object detector which uses a SOTA still image object detector and show that the class-specific optimized thresholds increase the accuracy against the common approach of using a general threshold for all classes. At https://github.com/cancam/LRP we provide the source code that can compute LRP for the PASCAL VOC and MSCOCO datasets. Our source code can easily be adapted to other datasets as well.

Önder Tuzcuoğlu, Aybora Köksal, Buğra Sofu et al.

We introduce, XoFTR, a cross-modal cross-view method for local feature matching between thermal infrared (TIR) and visible images. Unlike visible images, TIR images are less susceptible to adverse lighting and weather conditions but present difficulties in matching due to significant texture and intensity differences. Current hand-crafted and learning-based methods for visible-TIR matching fall short in handling viewpoint, scale, and texture diversities. To address this, XoFTR incorporates masked image modeling pre-training and fine-tuning with pseudo-thermal image augmentation to handle the modality differences. Additionally, we introduce a refined matching pipeline that adjusts for scale discrepancies and enhances match reliability through sub-pixel level refinement. To validate our approach, we collect a comprehensive visible-thermal dataset, and show that our method outperforms existing methods on many benchmarks.

Selim Kuzucu, Jiaee Cheong, Hatice Gunes et al.

Unfair predictions of machine learning (ML) models impede their broad acceptance in real-world settings. Tackling this arduous challenge first necessitates defining what it means for an ML model to be fair. This has been addressed by the ML community with various measures of fairness that depend on the prediction outcomes of the ML models, either at the group level or the individual level. These fairness measures are limited in that they utilize point predictions, neglecting their variances, or uncertainties, making them susceptible to noise, missingness and shifts in data. In this paper, we first show that an ML model may appear to be fair with existing point-based fairness measures but biased against a demographic group in terms of prediction uncertainties. Then, we introduce new fairness measures based on different types of uncertainties, namely, aleatoric uncertainty and epistemic uncertainty. We demonstrate on many datasets that (i) our uncertainty-based measures are complementary to existing measures of fairness, and (ii) they provide more insights about the underlying issues leading to bias.

Jiaee Cheong, Aditya Bangar, Sinan Kalkan et al.

Machine learning bias in mental health is becoming an increasingly pertinent challenge. Despite promising efforts indicating that multitask approaches often work better than unitask approaches, there is minimal work investigating the impact of multitask learning on performance and fairness in depression detection nor leveraged it to achieve fairer prediction outcomes. In this work, we undertake a systematic investigation of using a multitask approach to improve performance and fairness for depression detection. We propose a novel gender-based task-reweighting method using uncertainty grounded in how the PHQ-8 questionnaire is structured. Our results indicate that, although a multitask approach improves performance and fairness compared to a unitask approach, the results are not always consistent and we see evidence of negative transfer and a reduction in the Pareto frontier, which is concerning given the high-stake healthcare setting. Our proposed approach of gender-based reweighting with uncertainty improves performance and fairness and alleviates both challenges to a certain extent. Our findings on each PHQ-8 subitem task difficulty are also in agreement with the largest study conducted on the PHQ-8 subitem discrimination capacity, thus providing the very first tangible evidence linking ML findings with large-scale empirical population studies conducted on the PHQ-8.

Angus Man Ho Kwok, Jiaee Cheong, Sinan Kalkan et al.

This paper presents the very first attempt to evaluate machine learning fairness for depression detection using electroencephalogram (EEG) data. We conduct experiments using different deep learning architectures such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Unit (GRU) networks across three EEG datasets: Mumtaz, MODMA and Rest. We employ five different bias mitigation strategies at the pre-, in- and post-processing stages and evaluate their effectiveness. Our experimental results show that bias exists in existing EEG datasets and algorithms for depression detection, and different bias mitigation methods address bias at different levels across different fairness measures.

Aytaç Sekmen, Fatih Emre Gunes, Furkan Horoz et al.

Monocular Depth Estimation (MDE) is crucial for autonomous lunar rover navigation using electro-optical cameras. However, deploying terrestrial MDE networks to the Moon brings a severe domain gap due to harsh shadows, textureless regolith, and zero atmospheric scattering. Existing evaluations rely on analogs that fail to replicate these conditions and lack actual metric ground truth. To address this, we present LuMon, a comprehensive benchmarking framework to evaluate MDE methods for lunar exploration. We introduce novel datasets featuring high-quality stereo ground truth depth from the real Chang'e-3 mission and the CHERI dark analog dataset. Utilizing this framework, we conduct a systematic zero-shot evaluation of state-of-the-art architectures across synthetic, analog, and real datasets. We rigorously assess performance against mission critical challenges like craters, rocks, extreme shading, and varying depth ranges. Furthermore, we establish a sim-to-real domain adaptation baseline by fine tuning a foundation model on synthetic data. While this adaptation yields drastic in-domain performance gains, it exhibits minimal generalization to authentic lunar imagery, highlighting a persistent cross-domain transfer gap. Our extensive analysis reveals the inherent limitations of current networks and sets a standard foundation to guide future advancements in extraterrestrial perception and domain adaptation.

Jiaee Cheong, Abtin Mogharabin, Paul Liang et al.

Early efforts on leveraging self-supervised learning (SSL) to improve machine learning (ML) fairness has proven promising. However, such an approach has yet to be explored within a multimodal context. Prior work has shown that, within a multimodal setting, different modalities contain modality-unique information that can complement information of other modalities. Leveraging on this, we propose a novel subject-level loss function to learn fairer representations via the following three mechanisms, adapting the variance-invariance-covariance regularization (VICReg) method: (i) the variance term, which reduces reliance on the protected attribute as a trivial solution; (ii) the invariance term, which ensures consistent predictions for similar individuals; and (iii) the covariance term, which minimizes correlational dependence on the protected attribute. Consequently, our loss function, coined as FAIRWELL, aims to obtain subject-independent representations, enforcing fairness in multimodal prediction tasks. We evaluate our method on three challenging real-world heterogeneous healthcare datasets (i.e. D-Vlog, MIMIC and MODMA) which contain different modalities of varying length and different prediction tasks. Our findings indicate that our framework improves overall fairness performance with minimal reduction in classification performance and significantly improves on the performance-fairness Pareto frontier.

Hazal Mogultay Ozcan, Sinan Kalkan, Fatos T. Yarman-Vural

In this paper, we propose a novel model called Learnable VAE (L-VAE), which learns a disentangled representation together with the hyperparameters of the cost function. L-VAE can be considered as an extension of \b{eta}-VAE, wherein the hyperparameter, \b{eta}, is empirically adjusted. L-VAE mitigates the limitations of \b{eta}-VAE by learning the relative weights of the terms in the loss function to control the dynamic trade-off between disentanglement and reconstruction losses. In the proposed model, the weight of the loss terms and the parameters of the model architecture are learned concurrently. An additional regularization term is added to the loss function to prevent bias towards either reconstruction or disentanglement losses. Experimental analyses show that the proposed L-VAE finds an effective balance between reconstruction fidelity and disentangling the latent dimensions. Comparisons of the proposed L-VAE against \b{eta}-VAE, VAE, ControlVAE, DynamicVAE, and σ-VAE on datasets, such as dSprites, MPI3D-complex, Falcor3D, and Isaac3D reveals that L-VAE consistently provides the best or the second best performances measured by a set of disentanglement metrics. Moreover, qualitative experiments on CelebA dataset, confirm the success of the L-VAE model for disentangling the facial attributes.

Yusuf Meric Karadag, Sinan Kalkan, Ipek Gursel Dino

The problem of Time-series Forecasting is generally addressed by recurrent, Transformer-based and the recently proposed Mamba-based architectures. However, existing architectures generally process their input at a single temporal scale, which may be sub-optimal for many tasks where information changes over multiple time scales. In this paper, we introduce a novel architecture called Multi-scale Mamba (ms-Mamba) to address this gap. ms-Mamba incorporates multiple temporal scales by using multiple Mamba blocks with different sampling rates ($Δ$s). Our experiments on many benchmarks demonstrate that ms-Mamba outperforms state-of-the-art approaches, including the recently proposed Transformer-based and Mamba-based models.

Osman Tursun, Sinan Kalkan, Simon Denman et al.

Zero-shot Composed Image Retrieval (ZS-CIR) enables image search using a reference image and a text prompt without requiring specialized text-image composition networks trained on large-scale paired data. However, current ZS-CIR approaches suffer from three critical limitations in their reliance on composed text embeddings: static query embedding representations, insufficient utilization of image embeddings, and suboptimal performance when fusing text and image embeddings. To address these challenges, we introduce the \textbf{Prompt Directional Vector (PDV)}, a simple yet effective training-free enhancement that captures semantic modifications induced by user prompts. PDV enables three key improvements: (1) Dynamic composed text embeddings where prompt adjustments are controllable via a scaling factor, (2) composed image embeddings through semantic transfer from text prompts to image features, and (3) weighted fusion of composed text and image embeddings that enhances retrieval by balancing visual and semantic similarity. Our approach serves as a plug-and-play enhancement for existing ZS-CIR methods with minimal computational overhead. Extensive experiments across multiple benchmarks demonstrate that PDV consistently improves retrieval performance when integrated with state-of-the-art ZS-CIR approaches, particularly for methods that generate accurate compositional embeddings. The code will be released upon publication.

Arda Alniak, Sinan Kalkan, Mustafa Mert Ankarali et al.

Traditional monocular Visual-Inertial Odometry (VIO) systems struggle in low-texture environments where sparse visual features are insufficient for accurate pose estimation. To address this, dense Monocular Depth Estimation (MDE) has been widely explored as a complementary information source. While recent Vision Transformer (ViT) based complex foundational models offer dense, geometrically consistent depth, their computational demands typically preclude them from real-time edge deployment. Our work bridges this gap by integrating learned depth priors directly into the VINS-Mono optimization backend. We propose a novel framework that enforces affine-invariant depth consistency and pairwise ordinal constraints, explicitly filtering unstable artifacts via variance-based gating. This approach strictly adheres to the computational limits of edge devices while robustly recovering metric scale. Extensive experiments on the TartanGround and M3ED datasets demonstrate that our method prevents divergence in challenging scenarios and delivers significant accuracy gains, reducing Absolute Trajectory Error (ATE) by up to 28.3%. Code will be made available.

Muhammed Can Keles, Bedrettin Cetinkaya, Sinan Kalkan et al.

While heatmap-based human pose estimation methods have shown strong performance, they suffer from three main problems: (P1) "Commonly used Mean Squared Error (MSE)" Loss may not always improve joint localization because it penalizes all pixel deviations equally, without focusing explicitly on sharpening and correctly localizing the peak corresponding to the joint; (P2) heatmaps are spatially and class-wise imbalanced; and, (P3) there is a discrepancy between the evaluation metric (i.e., mAP) and the loss functions. We propose ranking-based losses to address these issues. Both theoretically and empirically, we show that our proposed losses are superior to commonly used heatmap losses (MSE, KL-Divergence). Our losses considerably increase the correlation between confidence scores and localization qualities, which is desirable because higher correlation leads to more accurate instance selection during Non-Maximum Suppression (NMS) and better Average Precision (mAP) performance. We refer to the models trained with our losses as RSPose. We show the effectiveness of RSPose across two different modes: one-dimensional and two-dimensional heatmaps, on three different datasets (COCO, CrowdPose, MPII). To the best of our knowledge, we are the first to propose losses that align with the evaluation metric (mAP) for human pose estimation. RSPose outperforms the previous state of the art on the COCO-val set and achieves an mAP score of 79.9 with ViTPose-H, a vision transformer model for human pose estimation. We also improve SimCC Resnet-50, a coordinate classification-based pose estimation method, by 1.5 AP on the COCO-val set, achieving 73.6 AP.

Osman Tursun, Sinan Kalkan, Simon Denman et al.

Heatmaps have been instrumental in helping understand deep network decisions, and are a common approach for Explainable AI (XAI). While significant progress has been made in enhancing the informativeness and accessibility of heatmaps, heatmap analysis is typically very subjective and limited to domain experts. As such, developing automatic, scalable, and numerical analysis methods to make heatmap-based XAI more objective, end-user friendly, and cost-effective is vital. In addition, there is a need for comprehensive evaluation metrics to assess heatmap quality at a granular level.

Barış Can Çam, Kemal Öksüz, Fehmi Kahraman et al.

This paper introduces Generalized Mask-aware Intersection-over-Union (GmaIoU) as a new measure for positive-negative assignment of anchor boxes during training of instance segmentation methods. Unlike conventional IoU measure or its variants, which only consider the proximity of anchor and ground-truth boxes; GmaIoU additionally takes into account the segmentation mask. This enables GmaIoU to provide more accurate supervision during training. We demonstrate the effectiveness of GmaIoU by replacing IoU with our GmaIoU in ATSS, a state-of-the-art (SOTA) assigner. Then, we train YOLACT, a real-time instance segmentation method, using our GmaIoU-based ATSS assigner. The resulting YOLACT based on the GmaIoU assigner outperforms (i) ATSS with IoU by $\sim 1.0-1.5$ mask AP, (ii) YOLACT with a fixed IoU threshold assigner by $\sim 1.5-2$ mask AP over different image sizes and (iii) decreases the inference time by $25 \%$ owing to using less anchors. Taking advantage of this efficiency, we further devise GmaYOLACT, a faster and $+7$ mask AP points more accurate detector than YOLACT. Our best model achieves $38.7$ mask AP at $26$ fps on COCO test-dev establishing a new state-of-the-art for real-time instance segmentation.

Nikhil Churamani, Sinan Kalkan, Hatice Gunes

Most state-of-the-art approaches for Facial Action Unit (AU) detection rely upon evaluating facial expressions from static frames, encoding a snapshot of heightened facial activity. In real-world interactions, however, facial expressions are usually more subtle and evolve in a temporal manner requiring AU detection models to learn spatial as well as temporal information. In this paper, we focus on both spatial and spatio-temporal features encoding the temporal evolution of facial AU activation. For this purpose, we propose the Action Unit Lifecycle-Aware Capsule Network (AULA-Caps) that performs AU detection using both frame and sequence-level features. While at the frame-level the capsule layers of AULA-Caps learn spatial feature primitives to determine AU activations, at the sequence-level, it learns temporal dependencies between contiguous frames by focusing on relevant spatio-temporal segments in the sequence. The learnt feature capsules are routed together such that the model learns to selectively focus more on spatial or spatio-temporal information depending upon the AU lifecycle. The proposed model is evaluated on the commonly used BP4D and GFT benchmark datasets obtaining state-of-the-art results on both the datasets.

Faisal Alamri, Sinan Kalkan, Nicolas Pugeault

Deep neural network approaches have demonstrated high performance in object recognition (CNN) and detection (Faster-RCNN) tasks, but experiments have shown that such architectures are vulnerable to adversarial attacks (FFF, UAP): low amplitude perturbations, barely perceptible by the human eye, can lead to a drastic reduction in labeling performance. This article proposes a new context module, called \textit{Transformer-Encoder Detector Module}, that can be applied to an object detector to (i) improve the labeling of object instances; and (ii) improve the detector's robustness to adversarial attacks. The proposed model achieves higher mAP, F1 scores and AUC average score of up to 13\% compared to the baseline Faster-RCNN detector, and an mAP score 8 points higher on images subjected to FFF or UAP attacks due to the inclusion of both contextual and visual features extracted from scene and encoded into the model. The result demonstrates that a simple ad-hoc context module can improve the reliability of object detectors significantly.

M. Esat Kalfaoglu, Sinan Kalkan, A. Aydin Alatan

In this work, we combine 3D convolution with late temporal modeling for action recognition. For this aim, we replace the conventional Temporal Global Average Pooling (TGAP) layer at the end of 3D convolutional architecture with the Bidirectional Encoder Representations from Transformers (BERT) layer in order to better utilize the temporal information with BERT's attention mechanism. We show that this replacement improves the performances of many popular 3D convolution architectures for action recognition, including ResNeXt, I3D, SlowFast and R(2+1)D. Moreover, we provide the-state-of-the-art results on both HMDB51 and UCF101 datasets with 85.10% and 98.69% top-1 accuracy, respectively. The code is publicly available.

Jonas Tjomsland, Sinan Kalkan, Hatice Gunes

To date, endowing robots with an ability to assess social appropriateness of their actions has not been possible. This has been mainly due to (i) the lack of relevant and labelled data, and (ii) the lack of formulations of this as a lifelong learning problem. In this paper, we address these two issues. We first introduce the Socially Appropriate Domestic Robot Actions dataset (MANNERS-DB), which contains appropriateness labels of robot actions annotated by humans. To be able to control but vary the configurations of the scenes and the social settings, MANNERS-DB has been created utilising a simulation environment by uniformly sampling relevant contextual attributes. Secondly, we train and evaluate a baseline Bayesian Neural Network (BNN) that estimates social appropriateness of actions in the MANNERS-DB. Finally, we formulate learning social appropriateness of actions as a continual learning problem using the uncertainty of the BNN parameters. The experimental results show that the social appropriateness of robot actions can be predicted with a satisfactory level of precision. Our work takes robots one step closer to a human-like understanding of (social) appropriateness of actions, with respect to the social context they operate in. To facilitate reproducibility and further progress in this area, the MANNERS-DB, the trained models and the relevant code will be made publicly available.

Tian Xu, Jennifer White, Sinan Kalkan et al.

Recognition of expressions of emotions and affect from facial images is a well-studied research problem in the fields of affective computing and computer vision with a large number of datasets available containing facial images and corresponding expression labels. However, virtually none of these datasets have been acquired with consideration of fair distribution across the human population. Therefore, in this work, we undertake a systematic investigation of bias and fairness in facial expression recognition by comparing three different approaches, namely a baseline, an attribute-aware and a disentangled approach, on two well-known datasets, RAF-DB and CelebA. Our results indicate that: (i) data augmentation improves the accuracy of the baseline model, but this alone is unable to mitigate the bias effect; (ii) both the attribute-aware and the disentangled approaches fortified with data augmentation perform better than the baseline approach in terms of accuracy and fairness; (iii) the disentangled approach is the best for mitigating demographic bias; and (iv) the bias mitigation strategies are more suitable in the existence of uneven attribute distribution or imbalanced number of subgroup data.

Fatih Can Kurnaz, Burak Hocaoğlu, Mert Kaan Yılmaz et al.

Robots collaborating with humans in realistic environments will need to be able to detect the tools that can be used and manipulated. However, there is no available dataset or study that addresses this challenge in real settings. In this paper, we fill this gap by providing an extensive dataset (METU-ALET) for detecting farming, gardening, office, stonemasonry, vehicle, woodworking and workshop tools. The scenes correspond to sophisticated environments with or without humans using the tools. The scenes we consider introduce several challenges for object detection, including the small scale of the tools, their articulated nature, occlusion, inter-class invariance, etc. Moreover, we train and compare several state of the art deep object detectors (including Faster R-CNN, Cascade R-CNN, RepPoint and RetinaNet) on our dataset. We observe that the detectors have difficulty in detecting especially small-scale tools or tools that are visually similar to parts of other tools. This in turn supports the importance of our dataset and paper. With the dataset, the code and the trained models, our work provides a basis for further research into tools and their use in robotics applications.

Fethiye Irmak Doğan, Sinan Kalkan, Iolanda Leite

Referring to objects in a natural and unambiguous manner is crucial for effective human-robot interaction. Previous research on learning-based referring expressions has focused primarily on comprehension tasks, while generating referring expressions is still mostly limited to rule-based methods. In this work, we propose a two-stage approach that relies on deep learning for estimating spatial relations to describe an object naturally and unambiguously with a referring expression. We compare our method to the state of the art algorithm in ambiguous environments (e.g., environments that include very similar objects with similar relationships). We show that our method generates referring expressions that people find to be more accurate ($\sim$30% better) and would prefer to use ($\sim$32% more often).

Ilker Bozcan, Sinan Kalkan

Scene modeling is very crucial for robots that need to perceive, reason about and manipulate the objects in their environments. In this paper, we adapt and extend Boltzmann Machines (BMs) for contextualized scene modeling. Although there are many models on the subject, ours is the first to bring together objects, relations, and affordances in a highly-capable generative model. For this end, we introduce a hybrid version of BMs where relations and affordances are introduced with shared, tri-way connections into the model. Moreover, we contribute a dataset for relation estimation and modeling studies. We evaluate our method in comparison with several baselines on object estimation, out-of-context object detection, relation estimation, and affordance estimation tasks. Moreover, to illustrate the generative capability of the model, we show several example scenes that the model is able to generate.

İlker Bozcan, Yağmur Oymak, İdil Zeynep Alemdar et al.

Scene models allow robots to reason about what is in the scene, what else should be in it, and what should not be in it. In this paper, we propose a hybrid Boltzmann Machine (BM) for scene modeling where relations between objects are integrated. To be able to do that, we extend BM to include tri-way edges between visible (object) nodes and make the network to share the relations across different objects. We evaluate our method against several baseline models (Deep Boltzmann Machines, and Restricted Boltzmann Machines) on a scene classification dataset, and show that it performs better in several scene reasoning tasks.

Fethiye Irmak Doğan, İlker Bozcan, Mehmet Çelik et al.

There have been several attempts at modeling context in robots. However, either these attempts assume a fixed number of contexts or use a rule-based approach to determine when to increment the number of contexts. In this paper, we pose the task of when to increment as a learning problem, which we solve using a Recurrent Neural Network. We show that the network successfully (with 98\% testing accuracy) learns to predict when to increment, and demonstrate, in a scene modeling problem (where the correct number of contexts is not known), that the robot increments the number of contexts in an expected manner (i.e., the entropy of the system is reduced). We also present how the incremental model can be used for various scene reasoning tasks.

Fethiye Irmak Doğan, Hande Çelikkanat, Sinan Kalkan

Context is an essential capability for robots that are to be as adaptive as possible in challenging environments. Although there are many context modeling efforts, they assume a fixed structure and number of contexts. In this paper, we propose an incremental deep model that extends Restricted Boltzmann Machines. Our model gets one scene at a time, and gradually extends the contextual model when necessary, either by adding a new context or a new context layer to form a hierarchy. We show on a scene classification benchmark that our method converges to a good estimate of the contexts of the scenes, and performs better or on-par on several tasks compared to other incremental models or non-incremental models.

Cemal Aker, Sinan Kalkan

Drone detection is the problem of finding the smallest rectangle that encloses the drone(s) in a video sequence. In this study, we propose a solution using an end-to-end object detection model based on convolutional neural networks. To solve the scarce data problem for training the network, we propose an algorithm for creating an extensive artificial dataset by combining background-subtracted real images. With this approach, we can achieve precision and recall values both of which are high at the same time.

Osman Tursun, Cemal Aker, Sinan Kalkan

Trademark retrieval (TR) has become an important yet challenging problem due to an ever increasing trend in trademark applications and infringement incidents. There have been many promising attempts for the TR problem, which, however, fell impracticable since they were evaluated with limited and mostly trivial datasets. In this paper, we provide a large-scale dataset with benchmark queries with which different TR approaches can be evaluated systematically. Moreover, we provide a baseline on this benchmark using the widely-used methods applied to TR in the literature. Furthermore, we identify and correct two important issues in TR approaches that were not addressed before: reversal of contrast, and presence of irrelevant text in trademarks severely affect the TR methods. Lastly, we applied deep learning, namely, several popular Convolutional Neural Network models, to the TR problem. To the best of the authors, this is the first attempt to do so.