Mehmet Kerim Yucel

Mehmet Kerim Yucel, Ramazan Gokberk Cinbis, Pinar Duygulu

Convolutional Neural Networks (CNN) are known to exhibit poor generalization performance under distribution shifts. Their generalization have been studied extensively, and one line of work approaches the problem from a frequency-centric perspective. These studies highlight the fact that humans and CNNs might focus on different frequency components of an image. First, inspired by these observations, we propose a simple yet effective data augmentation method HybridAugment that reduces the reliance of CNNs on high-frequency components, and thus improves their robustness while keeping their clean accuracy high. Second, we propose HybridAugment++, which is a hierarchical augmentation method that attempts to unify various frequency-spectrum augmentations. HybridAugment++ builds on HybridAugment, and also reduces the reliance of CNNs on the amplitude component of images, and promotes phase information instead. This unification results in competitive to or better than state-of-the-art results on clean accuracy (CIFAR-10/100 and ImageNet), corruption benchmarks (ImageNet-C, CIFAR-10-C and CIFAR-100-C), adversarial robustness on CIFAR-10 and out-of-distribution detection on various datasets. HybridAugment and HybridAugment++ are implemented in a few lines of code, does not require extra data, ensemble models or additional networks.

Roy Miles, Mehmet Kerim Yucel, Bruno Manganelli et al.

This paper tackles the problem of semi-supervised video object segmentation on resource-constrained devices, such as mobile phones. We formulate this problem as a distillation task, whereby we demonstrate that small space-time-memory networks with finite memory can achieve competitive results with state of the art, but at a fraction of the computational cost (32 milliseconds per frame on a Samsung Galaxy S22). Specifically, we provide a theoretically grounded framework that unifies knowledge distillation with supervised contrastive representation learning. These models are able to jointly benefit from both pixel-wise contrastive learning and distillation from a pre-trained teacher. We validate this loss by achieving competitive J&F to state of the art on both the standard DAVIS and YouTube benchmarks, despite running up to 5x faster, and with 32x fewer parameters.

Konstantinos Georgiadis, Albert Saà-Garriga, Mehmet Kerim Yucel et al.

Bokeh effect highlights an object (or any part of the image) while blurring the rest of the image, and creates a visually pleasant artistic effect. Due to the sensor-based limitations on mobile devices, machine learning (ML) based bokeh rendering has gained attention as a reliable alternative. In this paper, we focus on several improvements in ML-based bokeh rendering; i) on-device performance with high-resolution images, ii) ability to guide bokeh generation with user-editable masks and iii) ability to produce varying blur strength. To this end, we propose Adaptive Mask-based Pyramid Network (AMPN), which is formed of a Mask-Guided Bokeh Generator (MGBG) block and a Laplacian Pyramid Refinement (LPR) block. MGBG consists of two lightweight networks stacked to each other to generate the bokeh effect, and LPR refines and upsamples the output of MGBG to produce the high-resolution bokeh image. We achieve i) via our lightweight, mobile-friendly design choices, ii) via the stacked-network design of MGBG and the weakly-supervised mask prediction scheme and iii) via manually or automatically editing the intensity values of the mask that guide the bokeh generation. In addition to these features, our results show that AMPN produces competitive or better results compared to existing methods on the EBB! dataset, while being faster and smaller than the alternatives.

Evangelos Skartados, Konstantinos Georgiadis, Mehmet Kerim Yucel et al.

Space-time memory (STM) network methods have been dominant in semi-supervised video object segmentation (SVOS) due to their remarkable performance. In this work, we identify three key aspects where we can improve such methods; i) supervisory signal, ii) pretraining and iii) spatial awareness. We then propose TrickVOS; a generic, method-agnostic bag of tricks addressing each aspect with i) a structure-aware hybrid loss, ii) a simple decoder pretraining regime and iii) a cheap tracker that imposes spatial constraints in model predictions. Finally, we propose a lightweight network and show that when trained with TrickVOS, it achieves competitive results to state-of-the-art methods on DAVIS and YouTube benchmarks, while being one of the first STM-based SVOS methods that can run in real-time on a mobile device.

Evangelos Skartados, Mehmet Kerim Yucel, Bruno Manganelli et al.

Neural Radiance Fields (NeRF) have quickly become the primary approach for 3D reconstruction and novel view synthesis in recent years due to their remarkable performance. Despite the huge interest in NeRF methods, a practical use case of NeRFs has largely been ignored; the exploration of the scene space modelled by a NeRF. In this paper, for the first time in the literature, we propose and formally define the scene exploration framework as the efficient discovery of NeRF model inputs (i.e. coordinates and viewing angles), using which one can render novel views that adhere to user-selected criteria. To remedy the lack of approaches addressing scene exploration, we first propose two baseline methods called Guided-Random Search (GRS) and Pose Interpolation-based Search (PIBS). We then cast scene exploration as an optimization problem, and propose the criteria-agnostic Evolution-Guided Pose Search (EGPS) for efficient exploration. We test all three approaches with various criteria (e.g. saliency maximization, image quality maximization, photo-composition quality improvement) and show that our EGPS performs more favourably than other baselines. We finally highlight key points and limitations, and outline directions for future research in scene exploration.

Francesco Barbato, Umberto Michieli, Mehmet Kerim Yucel et al.

In multimedia understanding tasks, corrupted samples pose a critical challenge, because when fed to machine learning models they lead to performance degradation. In the past, three groups of approaches have been proposed to handle noisy data: i) enhancer and denoiser modules to improve the quality of the noisy data, ii) data augmentation approaches, and iii) domain adaptation strategies. All the aforementioned approaches come with drawbacks that limit their applicability; the first has high computational costs and requires pairs of clean-corrupted data for training, while the others only allow deployment of the same task/network they were trained on (\ie, when upstream and downstream task/network are the same). In this paper, we propose SyMPIE to solve these shortcomings. To this end, we design a small, modular, and efficient (just 2GFLOPs to process a Full HD image) system to enhance input data for robust downstream multimedia understanding with minimal computational cost. Our SyMPIE is pre-trained on an upstream task/network that should not match the downstream ones and does not need paired clean-corrupted samples. Our key insight is that most input corruptions found in real-world tasks can be modeled through global operations on color channels of images or spatial filters with small kernels. We validate our approach on multiple datasets and tasks, such as image classification (on ImageNetC, ImageNetC-Bar, VizWiz, and a newly proposed mixed corruption benchmark named ImageNetC-mixed) and semantic segmentation (on Cityscapes, ACDC, and DarkZurich) with consistent improvements of about 5\% relative accuracy gain across the board. The code of our approach and the new ImageNetC-mixed benchmark will be made available upon publication.

Anil Armagan, Albert Saà-Garriga, Bruno Manganelli et al.

Gaussian splatting (GS) for 3D reconstruction has become quite popular due to their fast training, inference speeds and high quality reconstruction. However, GS-based reconstructions generally consist of millions of Gaussians, which makes them hard to use on computationally constrained devices such as smartphones. In this paper, we first propose a principled analysis of advances in efficient GS methods. Then, we propose Trick-GS, which is a careful combination of several strategies including (1) progressive training with resolution, noise and Gaussian scales, (2) learning to prune and mask primitives and SH bands by their significance, and (3) accelerated GS training framework. Trick-GS takes a large step towards resource-constrained GS, where faster run-time, smaller and faster-convergence of models is of paramount concern. Our results on three datasets show that Trick-GS achieves up to 2x faster training, 40x smaller disk size and 2x faster rendering speed compared to vanilla GS, while having comparable accuracy.

Konstantinos Georgiadis, Mehmet Kerim Yucel, Albert Saa-Garriga

Single-view novel view synthesis (NVS) is a notorious problem due to its ill-posed nature, and often requires large, computationally expensive approaches to produce tangible results. In this paper, we propose CheapNVS: a fully end-to-end approach for narrow baseline single-view NVS based on a novel, efficient multiple encoder/decoder design trained in a multi-stage fashion. CheapNVS first approximates the laborious 3D image warping with lightweight learnable modules that are conditioned on the camera pose embeddings of the target view, and then performs inpainting on the occluded regions in parallel to achieve significant performance gains. Once trained on a subset of Open Images dataset, CheapNVS outperforms the state-of-the-art despite being 10 times faster and consuming 6% less memory. Furthermore, CheapNVS runs comfortably in real-time on mobile devices, reaching over 30 FPS on a Samsung Tab 9+.

Frederik Laboyrie, Mehmet Kerim Yucel, Albert Saa-Garriga

Dense prediction tasks have enjoyed a growing complexity of encoder architectures, decoders, however, have remained largely the same. They rely on individual blocks decoding intermediate feature maps sequentially. We introduce banks, shared structures that are used by each decoding block to provide additional context in the decoding process. These structures, through applying them via resampling and feature fusion, improve performance on depth estimation for state-of-the-art transformer-based architectures on natural and synthetic images whilst training on large-scale datasets.

Mehmet Kerim Yucel, Ramazan Gokberk Cinbis, Pinar Duygulu

Data shift robustness has been primarily investigated from a fully supervised perspective, and robustness of zero-shot learning (ZSL) models have been largely neglected. In this paper, we present novel analyses on the robustness of discriminative ZSL to image corruptions. We subject several ZSL models to a large set of common corruptions and defenses. In order to realize the corruption analysis, we curate and release the first ZSL corruption robustness datasets SUN-C, CUB-C and AWA2-C. We analyse our results by taking into account the dataset characteristics, class imbalance, class transitions between seen and unseen classes and the discrepancies between ZSL and GZSL performances. Our results show that discriminative ZSL suffers from corruptions and this trend is further exacerbated by the severe class imbalance and model weakness inherent in ZSL methods. We then combine our findings with those based on adversarial attacks in ZSL, and highlight the different effects of corruptions and adversarial examples, such as the pseudo-robustness effect present under adversarial attacks. We also obtain new strong baselines for both models with the defense methods. Finally, our experiments show that although existing methods to improve robustness somewhat work for ZSL models, they do not produce a tangible effect.

Mehmet Kerim Yucel, Valia Dimaridou, Anastasios Drosou et al.

Monocular (relative or metric) depth estimation is a critical task for various applications, such as autonomous vehicles, augmented reality and image editing. In recent years, with the increasing availability of mobile devices, accurate and mobile-friendly depth models have gained importance. Increasingly accurate models typically require more computational resources, which inhibits the use of such models on mobile devices. The mobile use case is arguably the most unrestricted one, which requires highly accurate yet mobile-friendly architectures. Therefore, we try to answer the following question: How can we improve a model without adding further complexity (i.e. parameters)? Towards this end, we systematically explore the design space of a relative depth estimation model from various dimensions and we show, with key design choices and ablation studies, even an existing architecture can reach highly competitive performance to the state of the art, with a fraction of the complexity. Our study spans an in-depth backbone model selection process, knowledge distillation, intermediate predictions, model pruning and loss rebalancing. We show that our model, using only DIW as the supervisory dataset, achieves 0.1156 WHDR on DIW with 2.6M parameters and reaches 37 FPS on a mobile GPU, without pruning or hardware-specific optimization. A pruned version of our model achieves 0.1208 WHDR on DIW with 1M parameters and reaches 44 FPS on a mobile GPU.

Yunus Can Bilge, Mehmet Kerim Yucel, Ramazan Gokberk Cinbis et al.

In many real-world problems, there is typically a large discrepancy between the characteristics of data used in training versus deployment. A prime example is the analysis of aggression videos: in a criminal incidence, typically suspects need to be identified based on their clean portrait-like photos, instead of their prior video recordings. This results in three major challenges; large domain discrepancy between violence videos and ID-photos, the lack of video examples for most individuals and limited training data availability. To mimic such scenarios, we formulate a realistic domain-transfer problem, where the goal is to transfer the recognition model trained on clean posed images to the target domain of violent videos, where training videos are available only for a subset of subjects. To this end, we introduce the WildestFaces dataset, tailored to study cross-domain recognition under a variety of adverse conditions. We divide the task of transferring a recognition model from the domain of clean images to the violent videos into two sub-problems and tackle them using (i) stacked affine-transforms for classifier-transfer, (ii) attention-driven pooling for temporal-adaptation. We additionally formulate a self-attention based model for domain-transfer. We establish a rigorous evaluation protocol for this clean-to-violent recognition task, and present a detailed analysis of the proposed dataset and the methods. Our experiments highlight the unique challenges introduced by the WildestFaces dataset and the advantages of the proposed approach.

Mehmet Kerim Yucel, Ramazan Gokberk Cinbis, Pinar Duygulu

Machine learning (ML) systems have introduced significant advances in various fields, due to the introduction of highly complex models. Despite their success, it has been shown multiple times that machine learning models are prone to imperceptible perturbations that can severely degrade their accuracy. So far, existing studies have primarily focused on models where supervision across all classes were available. In constrast, Zero-shot Learning (ZSL) and Generalized Zero-shot Learning (GZSL) tasks inherently lack supervision across all classes. In this paper, we present a study aimed on evaluating the adversarial robustness of ZSL and GZSL models. We leverage the well-established label embedding model and subject it to a set of established adversarial attacks and defenses across multiple datasets. In addition to creating possibly the first benchmark on adversarial robustness of ZSL models, we also present analyses on important points that require attention for better interpretation of ZSL robustness results. We hope these points, along with the benchmark, will help researchers establish a better understanding what challenges lie ahead and help guide their work.

Mehmet Kerim Yucel, Yunus Can Bilge, Oguzhan Oguz et al.

With the introduction of large-scale datasets and deep learning models capable of learning complex representations, impressive advances have emerged in face detection and recognition tasks. Despite such advances, existing datasets do not capture the difficulty of face recognition in the wildest scenarios, such as hostile disputes or fights. Furthermore, existing datasets do not represent completely unconstrained cases of low resolution, high blur and large pose/occlusion variances. To this end, we introduce the Wildest Faces dataset, which focuses on such adverse effects through violent scenes. The dataset consists of an extensive set of violent scenes of celebrities from movies. Our experimental results demonstrate that state-of-the-art techniques are not well-suited for violent scenes, and therefore, Wildest Faces is likely to stir further interest in face detection and recognition research.