İbrahim Batuhan Akkaya

Ibrahim Batuhan Akkaya, Kishaan Jeeveswaran, Bahram Zonooz et al.

The human visual system (HVS) employs foveated sampling and eye movements to achieve efficient perception, conserving both metabolic energy and computational resources. Drawing inspiration from this robustness and adaptability, we introduce the Foveated Dynamic Transformer (FDT), a foveation-guided dynamic token-selection architecture that integrates these mechanisms into a vision transformer framework. The FDT exhibits strong resilience to various types of noise and adversarial attacks, despite not being explicitly trained for such challenges. This inherent robustness is achieved through the use of fixation and foveation modules: the fixation module identifies fixation points to filter out irrelevant information, while the foveation module generates foveated embeddings with multi-scale information. At the 50% fixation-budget setting, FDT achieves higher accuracy than DeiT-S (81.9% vs. 80.9%) while reducing multiply-accumulate operations by 34.57%, highlighting one operating point on its accuracy-efficiency trade-off. These attributes position FDT as an HVS-inspired step toward artificial neural networks that combine adaptive computation with improved resilience.

Onur Can Koyun, Reyhan Kevser Keser, İbrahim Batuhan Akkaya et al.

Despite recent advances, object detection in aerial images is still a challenging task. Specific problems in aerial images makes the detection problem harder, such as small objects, densely packed objects, objects in different sizes and with different orientations. To address small object detection problem, we propose a two-stage object detection framework called "Focus-and-Detect". The first stage which consists of an object detector network supervised by a Gaussian Mixture Model, generates clusters of objects constituting the focused regions. The second stage, which is also an object detector network, predicts objects within the focal regions. Incomplete Box Suppression (IBS) method is also proposed to overcome the truncation effect of region search approach. Results indicate that the proposed two-stage framework achieves an AP score of 42.06 on VisDrone validation dataset, surpassing all other state-of-the-art small object detection methods reported in the literature, to the best of authors' knowledge.

Ibrahim Batuhan Akkaya, Ugur Halici

Unsupervised source-free domain adaptation methods aim to train a model for the target domain utilizing a pretrained source-domain model and unlabeled target-domain data, particularly when accessibility to source data is restricted due to intellectual property or privacy concerns. Traditional methods usually use self-training with pseudo-labeling, which is often subjected to thresholding based on prediction confidence. However, such thresholding limits the effectiveness of self-training due to insufficient supervision. This issue becomes more severe in a source-free setting, where supervision comes solely from the predictions of the pre-trained source model. In this study, we propose a novel approach by incorporating a mean-teacher model, wherein the student network is trained using all predictions from the teacher network. Instead of employing thresholding on predictions, we introduce a method to weight the gradients calculated from pseudo-labels based on the reliability of the teacher's predictions. To assess reliability, we introduce a novel approach using proxy-based metric learning. Our method is evaluated in synthetic-to-real and cross-city scenarios, demonstrating superior performance compared to existing state-of-the-art methods.

Ibrahim Batuhan Akkaya, Senthilkumar S. Kathiresan, Elahe Arani et al.

Vision transformers (ViTs) achieve remarkable performance on large datasets, but tend to perform worse than convolutional neural networks (CNNs) when trained from scratch on smaller datasets, possibly due to a lack of local inductive bias in the architecture. Recent studies have therefore added locality to the architecture and demonstrated that it can help ViTs achieve performance comparable to CNNs in the small-size dataset regime. Existing methods, however, are architecture-specific or have higher computational and memory costs. Thus, we propose a module called Local InFormation Enhancer (LIFE) that extracts patch-level local information and incorporates it into the embeddings used in the self-attention block of ViTs. Our proposed module is memory and computation efficient, as well as flexible enough to process auxiliary tokens such as the classification and distillation tokens. Empirical results show that the addition of the LIFE module improves the performance of ViTs on small image classification datasets. We further demonstrate how the effect can be extended to downstream tasks, such as object detection and semantic segmentation. In addition, we introduce a new visualization method, Dense Attention Roll-Out, specifically designed for dense prediction tasks, allowing the generation of class-specific attention maps utilizing the attention maps of all tokens.

Ibrahim Batuhan Akkaya, Fazil Altinel, Ugur Halici

Deep models trained on large-scale RGB image datasets have shown tremendous success. It is important to apply such deep models to real-world problems. However, these models suffer from a performance bottleneck under illumination changes. Thermal IR cameras are more robust against such changes, and thus can be very useful for the real-world problems. In order to investigate efficacy of combining feature-rich visible spectrum and thermal image modalities, we propose an unsupervised domain adaptation method which does not require RGB-to-thermal image pairs. We employ large-scale RGB dataset MS-COCO as source domain and thermal dataset FLIR ADAS as target domain to demonstrate results of our method. Although adversarial domain adaptation methods aim to align the distributions of source and target domains, simply aligning the distributions cannot guarantee perfect generalization to the target domain. To this end, we propose a self-training guided adversarial domain adaptation method to promote generalization capabilities of adversarial domain adaptation methods. To perform self-training, pseudo labels are assigned to the samples on the target thermal domain to learn more generalized representations for the target domain. Extensive experimental analyses show that our proposed method achieves better results than the state-of-the-art adversarial domain adaptation methods. The code and models are publicly available.

Ibrahim Batuhan Akkaya, Ugur Halici

Image-to-image translation (IIT) has made much progress recently with the development of adversarial learning. In most of the recent work, an adversarial loss is utilized to match the distributions of the translated and target image sets. However, this may create artifacts if two domains have different marginal distributions, for example, in uniform areas. In this work, we propose an unsupervised IIT method that preserves the uniform regions after the translation. The gradient adjustment loss, which is the L2 norm between the Sobel response of the target image and the adjusted Sobel response of the source images, is utilized. The proposed method is validated on the jellyfish-to-Haeckel dataset, which is prepared to demonstrate the mentioned problem, which contains images with different background distributions. We demonstrate that our method obtained a performance gain compared to the baseline method qualitatively and quantitatively, showing the effectiveness of the proposed method.