Seungkyu Lee

Chaoning Zhang, Dongshen Han, Yu Qiao et al.

Segment Anything Model (SAM) has attracted significant attention due to its impressive zero-shot transfer performance and high versatility for numerous vision applications (like image editing with fine-grained control). Many of such applications need to be run on resource-constraint edge devices, like mobile phones. In this work, we aim to make SAM mobile-friendly by replacing the heavyweight image encoder with a lightweight one. A naive way to train such a new SAM as in the original SAM paper leads to unsatisfactory performance, especially when limited training sources are available. We find that this is mainly caused by the coupled optimization of the image encoder and mask decoder, motivated by which we propose decoupled distillation. Concretely, we distill the knowledge from the heavy image encoder (ViT-H in the original SAM) to a lightweight image encoder, which can be automatically compatible with the mask decoder in the original SAM. The training can be completed on a single GPU within less than one day, and the resulting lightweight SAM is termed MobileSAM which is more than 60 times smaller yet performs on par with the original SAM. For inference speed, With a single GPU, MobileSAM runs around 10ms per image: 8ms on the image encoder and 4ms on the mask decoder. With superior performance, our MobileSAM is around 5 times faster than the concurrent FastSAM and 7 times smaller, making it more suitable for mobile applications. Moreover, we show that MobileSAM can run relatively smoothly on CPU. The code for our project is provided at \href{https://github.com/ChaoningZhang/MobileSAM}{\textcolor{red}{MobileSAM}}), with a demo showing that MobileSAM can run relatively smoothly on CPU.

Dongsheng Han, Chaoning Zhang, Yu Qiao et al.

Meta AI Research has recently released SAM (Segment Anything Model) which is trained on a large segmentation dataset of over 1 billion masks. As a foundation model in the field of computer vision, SAM (Segment Anything Model) has gained attention for its impressive performance in generic object segmentation. Despite its strong capability in a wide range of zero-shot transfer tasks, it remains unknown whether SAM can detect things in challenging setups like transparent objects. In this work, we perform an empirical evaluation of two glass-related challenging scenarios: mirror and transparent objects. We found that SAM often fails to detect the glass in both scenarios, which raises concern for deploying the SAM in safety-critical situations that have various forms of glass.

Dongshen Han, Seungkyu Lee, Chaoning Zhang et al.

Glass surfaces of transparent objects and mirrors are not able to be uniquely and explicitly characterized by their visual appearances because they contain the visual appearance of other reflected or transmitted surfaces as well. Detecting glass regions from a single-color image is a challenging task. Recent deep-learning approaches have paid attention to the description of glass surface boundary where the transition of visual appearances between glass and non-glass surfaces are observed. In this work, we analytically investigate how glass surface boundary helps to characterize glass objects. Inspired by prior semantic segmentation approaches with challenging image types such as X-ray or CT scans, we propose separated internal-external boundary attention modules that individually learn and selectively integrate visual characteristics of the inside and outside region of glass surface from a single color image. Our proposed method is evaluated on six public benchmarks comparing with state-of-the-art methods showing promising results.

Seunghwan Kim, Seungkyu Lee

Variational autoencoder (VAE) is an established generative model but is notorious for its blurriness. In this work, we investigate the blurry output problem of VAE and resolve it, exploiting the variance of Gaussian decoder and $β$ of beta-VAE. Specifically, we reveal that the indistinguishability of decoder variance and $β$ hinders appropriate analysis of the model by random likelihood value, and limits performance improvement by omitting the gain from $β$. To address the problem, we propose Beta-Sigma VAE (BS-VAE) that explicitly separates $β$ and decoder variance $σ^2_x$ in the model. Our method demonstrates not only superior performance in natural image synthesis but also controllable parameters and predictable analysis compared to conventional VAE. In our experimental evaluation, we employ the analysis of rate-distortion curve and proxy metrics on computer vision datasets. The code is available on https://github.com/overnap/BS-VAE

Heechan Yoon, Seungkyu Lee

Unlike opaque object, novel view synthesis of transparent object is a challenging task, because transparent object refracts light of background causing visual distortions on the transparent object surface along the viewpoint change. Recently introduced Neural Radiance Fields (NeRF) is a view synthesis method. Thanks to its remarkable performance improvement, lots of following applications based on NeRF in various topics have been developed. However, if an object with a different refractive index is included in a scene such as transparent object, NeRF shows limited performance because refracted light ray at the surface of the transparent object is not appropriately considered. To resolve the problem, we propose a NeRF-based method consisting of the following three steps: First, we reconstruct a three-dimensional shape of a transparent object using visual hull. Second, we simulate the refraction of the rays inside of the transparent object according to Snell's law. Last, we sample points through refracted rays and put them into NeRF. Experimental evaluation results demonstrate that our method addresses the limitation of conventional NeRF with transparent objects.

Minjae Oh, Sangjun Song, Seungkyu Lee et al.

Small reasoning models (SRMs) often overthink during tool use: they reach a correct tool-argument configuration, then continue reasoning and overwrite it with an incorrect final call. We diagnose overthinking via oracle rollouts that inject </think> at sentence boundaries. On the Berkeley Function Calling Leaderboard (BFCL), this oracle termination lifts average accuracy from 85.8\% to 94.2\% while reducing tokens by 80-94\%, revealing substantial recoverable headroom and potential redundant reasoning. While prior work on concise reasoning has largely targeted mathematics, tool reasoning remains underexplored. We adapt various early-termination baselines to tool use and introduce ThinkBrake, a training-free decoding heuristic. ThinkBrake monitors the log-probability margin between </think> and the current top token at sentence boundaries and triggers termination when this margin becomes small. Across BFCL's single turn, non-live and live splits, ThinkBrake preserves or improves accuracy while reducing tokens up to 25\%, outperforming various baselines.

Seungkyu Lee, Nalim Kim, Yohan Jo

Tool agents -- LLM-based systems that interact with external APIs -- offer a way to execute real-world tasks. However, as tasks become increasingly complex, these agents struggle to identify and call the correct APIs in the proper order. To tackle this problem, we investigate converting API documentation into a structured API graph that captures API dependencies and leveraging it for multi-tool queries that require compositional API calls. To support this, we introduce In-N-Out, the first expert-annotated dataset of API graphs built from two real-world API benchmarks and their documentation. Using In-N-Out significantly improves performance on both tool retrieval and multi-tool query generation, nearly doubling that of LLMs using documentation alone. Moreover, graphs generated by models fine-tuned on In-N-Out close 90% of this gap, showing that our dataset helps models learn to comprehend API documentation and parameter relationships. Our findings highlight the promise of using explicit API graphs for tool agents and the utility of In-N-Out as a valuable resource. We will release the dataset and code publicly.

Hyunsoo Song, Seungwhan Kim, Seungkyu Lee

Variational autoencoders (VAEs), one of the most widely used generative models, are known to suffer from posterior collapse, a phenomenon that reduces the diversity of generated samples. To avoid posterior collapse, many prior works have tried to control the influence of regularization loss. However, the trade-off between reconstruction and regularization is not satisfactory. For this reason, several methods have been proposed to guarantee latent identifiability, which is the key to avoiding posterior collapse. However, they require structural constraints on the network architecture. For further clarification, we define local posterior collapse to reflect the importance of individual sample points in the data space and to relax the network constraint. Then, we propose Latent Reconstruction(LR) loss, which is inspired by mathematical properties of injective and composite functions, to control posterior collapse without restriction to a specific architecture. We experimentally evaluate our approach, which controls posterior collapse on varied datasets such as MNIST, fashionMNIST, Omniglot, CelebA, and FFHQ.

Samra Irshad, Seungkyu Lee, Nassir Navab et al.

The presence of adversarial examples in the physical world poses significant challenges to the deployment of Deep Neural Networks in safety-critical applications such as autonomous driving. Most existing methods for crafting physical-world adversarial examples are ad-hoc, relying on temporary modifications like shadows, laser beams, or stickers that are tailored to specific scenarios. In this paper, we introduce a new class of physical-world adversarial examples, AdvWT, which draws inspiration from the naturally occurring phenomenon of `wear and tear', an inherent property of physical objects. Unlike manually crafted perturbations, `wear and tear' emerges organically over time due to environmental degradation, as seen in the gradual deterioration of outdoor signboards. To achieve this, AdvWT follows a two-step approach. First, a GAN-based, unsupervised image-to-image translation network is employed to model these naturally occurring damages, particularly in the context of outdoor signboards. The translation network encodes the characteristics of damaged signs into a latent `damage style code'. In the second step, we introduce adversarial perturbations into the style code, strategically optimizing its transformation process. This manipulation subtly alters the damage style representation, guiding the network to generate adversarial images where the appearance of damages remains perceptually realistic, while simultaneously ensuring their effectiveness in misleading neural networks. Through comprehensive experiments on two traffic sign datasets, we show that AdvWT effectively misleads DNNs in both digital and physical domains. AdvWT achieves an effective attack success rate, greater robustness, and a more natural appearance compared to existing physical-world adversarial examples. Additionally, integrating AdvWT into training enhances a model's generalizability to real-world damaged signs.

Seunghwan Kim, Sunha Park, Seungkyu Lee

Prefracture method is a practical implementation for real-time object destruction that is hardly achievable within performance constraints, but can produce unrealistic results due to its heuristic nature. To mitigate it, we approach the clustering of prefractured mesh generation as an unordered segmentation on point cloud data, and propose leveraging the deep neural network trained on a physics-based dataset. Our novel paradigm successfully predicts the structural weakness of object that have been limited, exhibiting ready-to-use results with remarkable quality.

Dongshen Han, Seungkyu Lee, Chaoning Zhang et al.

Single Image Reflection Removal (SIRR) in real-world images is a challenging task due to diverse image degradations occurring on the glass surface during light transmission and reflection. Many existing methods rely on specific prior assumptions to resolve the problem. In this paper, we propose a general reflection intensity prior that captures the intensity of the reflection phenomenon and demonstrate its effectiveness. To learn the reflection intensity prior, we introduce the Reflection Prior Extraction Network (RPEN). By segmenting images into regional patches, RPEN learns non-uniform reflection prior in an image. We propose Prior-based Reflection Removal Network (PRRN) using a simple transformer U-Net architecture that adapts reflection prior fed from RPEN. Experimental results on real-world benchmarks demonstrate the effectiveness of our approach achieving state-of-the-art accuracy in SIRR.

Sohee Kim, Seungkyu Lee

Continual learning is a concept of online learning with multiple sequential tasks. One of the critical barriers of continual learning is that a network should learn a new task keeping the knowledge of old tasks without access to any data of the old tasks. In this paper, we propose a neuron activation importance-based regularization method for stable continual learning regardless of the order of tasks. We conduct comprehensive experiments on existing benchmark data sets to evaluate not just the stability and plasticity of our method with improved classification accuracy also the robustness of the performance along the changes of task order.

Sanghun Kim, SeungKyu Lee

Due to the unstable nature of minimax game between generator and discriminator, improving the performance of GANs is a challenging task. Recent studies have shown that selected high-quality samples in training improve the performance of GANs. However, sampling approaches which discard samples show limitations in some aspects such as the speed of training and optimality of the networks. In this paper we propose unrealistic feature suppression (UFS) module that keeps high-quality features and suppresses unrealistic features. UFS module keeps the training stability of networks and improves the quality of generated images. We demonstrate the effectiveness of UFS module on various models such as WGAN-GP, SNGAN, and BigGAN. By using UFS module, we achieved better Frechet inception distance and inception score compared to various baseline models. We also visualize how effectively our UFS module suppresses unrealistic features through class activation maps.

Gahye Lee, Seungkyu Lee

Anomaly detection in data analysis is an interesting but still challenging research topic in real world applications. As the complexity of data dimension increases, it requires to understand the semantic contexts in its description for effective anomaly characterization. However, existing anomaly detection methods show limited performances with high dimensional data such as ImageNet. Existing studies have evaluated their performance on low dimensional, clean and well separated data set such as MNIST and CIFAR-10. In this paper, we study anomaly detection with high dimensional and complex normal data. Our observation is that, in general, anomaly data is defined by semantically explainable features which are able to be used in defining semantic sub-clusters of normal data as well. We hypothesize that if there exists reasonably good feature space semantically separating sub-clusters of given normal data, unseen anomaly also can be well distinguished in the space from the normal data. We propose to perform semantic clustering on given normal data and train a classifier to learn the discriminative feature space where anomaly detection is finally performed. Based on our careful and extensive experimental evaluations with MNIST, CIFAR-10, and ImageNet with various combinations of normal and anomaly data, we show that our anomaly detection scheme outperforms state of the art methods especially with high dimensional real world images.

Gahye Lee, Seungkyu Lee

Generative Adversarial Networks (GAN) are trained to generate sample images of interest distribution. To this end, generator network of GAN learns implicit distribution of real data set from the classification with candidate generated samples. Recently, various GANs have suggested novel ideas for stable optimizing of its networks. However, in real implementation, sometimes they still represent a only narrow part of true distribution or fail to converge. We assume this ill posed problem comes from poor gradient from objective function of discriminator, which easily trap the generator in a bad situation. To address this problem, we propose a mode penalty GAN combined with pre-trained auto encoder for explicit representation of generated and real data samples in the encoded space. In this space, we make a generator manifold to follow a real manifold by finding entire modes of target distribution. In addition, penalty for uncovered modes of target distribution is given to the generator which encourages it to find overall target distribution. We demonstrate that applying the proposed method to GANs helps generator's optimization becoming more stable and having faster convergence through experimental evaluations.

Sherzod Salokhiddinov, Seungkyu Lee

Face detection and tracking in low resolution image is not a trivial task due to the limitation in the appearance features for face characterization. Moreover, facial expression gives additional distortion on this small and noisy face. In this paper, we propose deformable symmetric Gabor wavelet network face model for face detection in low resolution image. Our model optimizes the rotation, translation, dilation, perspective and partial deformation amount of the face model with symmetry constraints. Symmetry constraints help our model to be more robust to noise and distortion. Experimental results on our low resolution face image dataset and videos show promising face detection and tracking results under various challenging conditions.