Sangmin Lee

Geonmo Gu, Byeongho Heo, Jaemyung Yu et al.

Universal Multimodal embedding models built on Multimodal Large Language Models (MLLMs) have traditionally employed contrastive learning, which aligns representations of query-target pairs across different modalities. Yet, despite its empirical success, they are primarily built on a "single-turn" formulation where each query-target pair is treated as an independent data point. This paradigm leads to computational inefficiency when scaling, as it requires a separate forward pass for each pair and overlooks potential contextual relationships between multiple queries that can relate to the same context. In this work, we introduce Multi-Turn Contrastive Learning (MuCo), a dialogue-inspired framework that revisits this process. MuCo leverages the conversational nature of MLLMs to process multiple, related query-target pairs associated with a single image within a single forward pass. This allows us to extract a set of multiple query and target embeddings simultaneously, conditioned on a shared context representation, amplifying the effective batch size and overall training efficiency. Experiments exhibit MuCo with a newly curated 5M multimodal multi-turn dataset (M3T), which yields state-of-the-art retrieval performance on MMEB and M-BEIR benchmarks, while markedly enhancing both training efficiency and representation coherence across modalities. Code and M3T are available at https://github.com/naver-ai/muco

Joonghoon Kim, Saeran Park, Kiyoon Jeong et al.

This paper describes the DSBA submissions to the Prompting Large Language Models as Explainable Metrics shared task, where systems were submitted to two tracks: small and large summarization tracks. With advanced Large Language Models (LLMs) such as GPT-4, evaluating the quality of Natural Language Generation (NLG) has become increasingly paramount. Traditional similarity-based metrics such as BLEU and ROUGE have shown to misalign with human evaluation and are ill-suited for open-ended generation tasks. To address this issue, we explore the potential capability of LLM-based metrics, especially leveraging open-source LLMs. In this study, wide range of prompts and prompting techniques are systematically analyzed with three approaches: prompting strategy, score aggregation, and explainability. Our research focuses on formulating effective prompt templates, determining the granularity of NLG quality scores and assessing the impact of in-context examples on LLM-based evaluation. Furthermore, three aggregation strategies are compared to identify the most reliable method for aggregating NLG quality scores. To examine explainability, we devise a strategy that generates rationales for the scores and analyzes the characteristics of the explanation produced by the open-source LLMs. Extensive experiments provide insights regarding evaluation capabilities of open-source LLMs and suggest effective prompting strategies.

Geon Yeong Park, Sangmin Lee, Sang Wan Lee et al.

Neural networks are often biased to spuriously correlated features that provide misleading statistical evidence that does not generalize. This raises an interesting question: ``Does an optimal unbiased functional subnetwork exist in a severely biased network? If so, how to extract such subnetwork?" While empirical evidence has been accumulated about the existence of such unbiased subnetworks, these observations are mainly based on the guidance of ground-truth unbiased samples. Thus, it is unexplored how to discover the optimal subnetworks with biased training datasets in practice. To address this, here we first present our theoretical insight that alerts potential limitations of existing algorithms in exploring unbiased subnetworks in the presence of strong spurious correlations. We then further elucidate the importance of bias-conflicting samples on structure learning. Motivated by these observations, we propose a Debiased Contrastive Weight Pruning (DCWP) algorithm, which probes unbiased subnetworks without expensive group annotations. Experimental results demonstrate that our approach significantly outperforms state-of-the-art debiasing methods despite its considerable reduction in the number of parameters.

Bikram Boote, Anh Thai, Wenqi Jia et al.

Point tracking is a fundamental problem in computer vision with numerous applications in AR and robotics. A common failure mode in long-term point tracking occurs when the predicted point leaves the object it belongs to and lands on the background or another object. We identify this as the failure to correctly capture objectness properties in learning to track. To address this limitation of prior work, we propose a novel objectness regularization approach that guides points to be aware of object priors by forcing them to stay inside the the boundaries of object instances. By capturing objectness cues at training time, we avoid the need to compute object masks during testing. In addition, we leverage contextual attention to enhance the feature representation for capturing objectness at the feature level more effectively. As a result, our approach achieves state-of-the-art performance on three point tracking benchmarks, and we further validate the effectiveness of our components via ablation studies. The source code is available at: https://github.com/RehgLab/tracking_objectness

Sangmin Lee, Byeongsu Sim, Jong Chul Ye

To understand learning the dynamics of deep ReLU networks, we investigate the dynamic system of gradient flow $w(t)$ by decomposing it to magnitude $w(t)$ and angle $φ(t):= π- θ(t) $ components. In particular, for multi-layer single ReLU neurons with spherically symmetric data distribution and the square loss function, we provide upper and lower bounds for magnitude and angle components to describe the dynamics of gradient flow. Using the obtained bounds, we conclude that small scale initialization induces slow convergence speed for deep single ReLU neurons. Finally, by exploiting the relation of gradient flow and gradient descent, we extend our results to the gradient descent approach. All theoretical results are verified by experiments.

Geon Yeong Park, Chanyong Jung, Sangmin Lee et al.

Spurious correlations can cause strong biases in deep neural networks, impairing generalization ability. While most existing debiasing methods require full supervision on either spurious attributes or target labels, training a debiased model from a limited amount of both annotations is still an open question. To address this issue, we investigate an interesting phenomenon using the spectral analysis of latent representations: spuriously correlated attributes make neural networks inductively biased towards encoding lower effective rank representations. We also show that a rank regularization can amplify this bias in a way that encourages highly correlated features. Leveraging these findings, we propose a self-supervised debiasing framework potentially compatible with unlabeled samples. Specifically, we first pretrain a biased encoder in a self-supervised manner with the rank regularization, serving as a semantic bottleneck to enforce the encoder to learn the spuriously correlated attributes. This biased encoder is then used to discover and upweight bias-conflicting samples in a downstream task, serving as a boosting to effectively debias the main model. Remarkably, the proposed debiasing framework significantly improves the generalization performance of self-supervised learning baselines and, in some cases, even outperforms state-of-the-art supervised debiasing approaches.

SiHun Lee, Sangmin Lee, Kijoo Jang et al.

A data-driven parametric model order reduction (MOR) method using a deep artificial neural network is proposed. The present network, which is the least-squares hierarchical variational autoencoder (LSH-VAE), is capable of performing nonlinear MOR for the parametric interpolation of a nonlinear dynamic system with a significant number of degrees of freedom. LSH-VAE exploits two major changes to the existing networks: a hierarchical deep structure and a hybrid weighted, probabilistic loss function. The enhancements result in a significantly improved accuracy and stability compared against the conventional nonlinear MOR methods, autoencoder, and variational autoencoder. Upon LSH-VAE, a parametric MOR framework is presented based on the spherically linear interpolation of the latent manifold. The present framework is validated and evaluated on three nonlinear and multiphysics dynamic systems. First, the present framework is evaluated on the fluid-structure interaction benchmark problem to assess its efficiency and accuracy. Then, a highly nonlinear aeroelastic phenomenon, limit cycle oscillation, is analyzed. Finally, the present framework is applied to a three-dimensional fluid flow to demonstrate its capability of efficiently analyzing a significantly large number of degrees of freedom. The performance of LSH-VAE is emphasized by comparing its results against that of the widely used nonlinear MOR methods, convolutional autoencoder, and $β$-VAE. The present framework exhibits a significantly enhanced accuracy to the conventional methods while still exhibiting a large speed-up factor.

Evren Çetinkaya, Sangmin Lee, Jung Uk Kim et al.

Visual prompting has emerged as a powerful method for adapting pre-trained models to new domains without updating model parameters. However, existing prompting methods typically optimize a single prompt per domain and apply it uniformly to all inputs, limiting their ability to generalize under intra and inter-domain variability, which is especially critical in the medical field. To address this, we propose APEX, an Adaptive Prompt EXtraction framework that retrieves input-specific prompts from a learnable prompt memory. The memory stores diverse, domain-discriminative prompt representations and is queried via domain features extracted from the Fourier spectrum. To learn robust and discriminative domain features, we introduce a novel Low-Frequency Feature Contrastive (LFC) learning framework that clusters representations from the same domain while separating those from different domains. Extensive experiments on two medical segmentation tasks demonstrate that APEX significantly improves generalization across both seen and unseen domains. Furthermore, it complements any existing backbones and consistently enhances performance, confirming its effectiveness as a plug-and-play prompting solution in medical fields. The code is available at https://github.com/cetinkayaevren/apex/

Chanhui Lee, Juhyeon Kim, Yongjun Jeong et al.

Scaling laws have allowed Pre-trained Language Models (PLMs) into the field of causal reasoning. Causal reasoning of PLM relies solely on text-based descriptions, in contrast to causal discovery which aims to determine the causal relationships between variables utilizing data. Recently, there has been current research regarding a method that mimics causal discovery by aggregating the outcomes of repetitive causal reasoning, achieved through specifically designed prompts. It highlights the usefulness of PLMs in discovering cause and effect, which is often limited by a lack of data, especially when dealing with multiple variables. Conversely, the characteristics of PLMs which are that PLMs do not analyze data and they are highly dependent on prompt design leads to a crucial limitation for directly using PLMs in causal discovery. Accordingly, PLM-based causal reasoning deeply depends on the prompt design and carries out the risk of overconfidence and false predictions in determining causal relationships. In this paper, we empirically demonstrate the aforementioned limitations of PLM-based causal reasoning through experiments on physics-inspired synthetic data. Then, we propose a new framework that integrates prior knowledge obtained from PLM with a causal discovery algorithm. This is accomplished by initializing an adjacency matrix for causal discovery and incorporating regularization using prior knowledge. Our proposed framework not only demonstrates improved performance through the integration of PLM and causal discovery but also suggests how to leverage PLM-extracted prior knowledge with existing causal discovery algorithms.

Joohyeok Kim, Joonhyeon Song, Seohwan Yun et al.

The Foundation model for image segmentation, Segment Anything (SAM), has been actively researched in various fields since its proposal. Various researches have been proposed to adapt SAM to specific domains, with one notable approach involving the addition and training of lightweight adapter modules. While adapter-based fine-tuning approaches have reported parameter efficiency and significant performance improvements, they face a often overlooked issue: the excessive consumption of GPU memory relative to the number of trainable parameters. Addressing this issue, this paper proposes a memory-efficient parallel convolutional adapter architecture. This architecture connects in parallel with SAM's image encoder, eliminating the need to store activations and gradients of the image encoder during model training. Our proposed architecture demonstrated competitive experimental results while using less than half the GPU memory compared to SAM Adapter, indicating its value as an alternative to simple decoder fine-tuning when hardware limitations preclude adapter-based learning. Our code implementation is available at our github.

Dongjin Kim, Sung Jin Um, Sangmin Lee et al.

The goal of the multi-sound source localization task is to localize sound sources from the mixture individually. While recent multi-sound source localization methods have shown improved performance, they face challenges due to their reliance on prior information about the number of objects to be separated. In this paper, to overcome this limitation, we present a novel multi-sound source localization method that can perform localization without prior knowledge of the number of sound sources. To achieve this goal, we propose an iterative object identification (IOI) module, which can recognize sound-making objects in an iterative manner. After finding the regions of sound-making objects, we devise object similarity-aware clustering (OSC) loss to guide the IOI module to effectively combine regions of the same object but also distinguish between different objects and backgrounds. It enables our method to perform accurate localization of sound-making objects without any prior knowledge. Extensive experimental results on the MUSIC and VGGSound benchmarks show the significant performance improvements of the proposed method over the existing methods for both single and multi-source. Our code is available at: https://github.com/VisualAIKHU/NoPrior_MultiSSL

Fiona Ryan, Ajay Bati, Sangmin Lee et al. · gatech

We address the problem of gaze target estimation, which aims to predict where a person is looking in a scene. Predicting a person's gaze target requires reasoning both about the person's appearance and the contents of the scene. Prior works have developed increasingly complex, hand-crafted pipelines for gaze target estimation that carefully fuse features from separate scene encoders, head encoders, and auxiliary models for signals like depth and pose. Motivated by the success of general-purpose feature extractors on a variety of visual tasks, we propose Gaze-LLE, a novel transformer framework that streamlines gaze target estimation by leveraging features from a frozen DINOv2 encoder. We extract a single feature representation for the scene, and apply a person-specific positional prompt to decode gaze with a lightweight module. We demonstrate state-of-the-art performance across several gaze benchmarks and provide extensive analysis to validate our design choices. Our code is available at: http://github.com/fkryan/gazelle .

Jeongsoo Park, Dong-Gyun Han, Hyoung Sul La et al.

This paper presents a novel deep learning approach for analyzing massive underwater acoustic data by leveraging a model trained on a broad spectrum of non-underwater (aerial) sounds. Recognizing the challenge in labeling vast amounts of underwater data, we propose a two-fold methodology to accelerate this labor-intensive procedure. The first part of our approach involves PCA and UMAP visualization of the underwater data using the feature vectors of an aerial sound recognition model. This enables us to cluster the data in a two dimensional space and listen to points within these clusters to understand their defining characteristics. This innovative method simplifies the process of selecting candidate labels for further training. In the second part, we train a neural network model using both the selected underwater data and the non-underwater dataset. We conducted a quantitative analysis to measure the precision, recall, and F1 score of our model for recognizing airgun sounds, a common type of underwater sound. The F1 score achieved by our model exceeded 84.3%, demonstrating the effectiveness of our approach in analyzing underwater acoustic data. The methodology presented in this paper holds significant potential to reduce the amount of labor required in underwater data analysis and opens up new possibilities for further research in the field of cross-domain data analysis.

Junho Kim, Xu Cao, Houze Yang et al.

Understanding social interactions requires reasoning over subtle non-verbal cues, yet current multimodal large language models (MLLMs) often fail to identify who interacts with whom in multi-person videos. We introduce GRASP, a large-scale social reasoning dataset that connects high-level social QA with fine-grained gaze and deictic gesture events. GRASP contains 290K question--answer pairs over 46K videos totaling 749 hours, organized by a 16-category taxonomy spanning gaze, gesture, and joint gaze--gesture reasoning, together with GRASP-Bench for evaluation. Unlike prior resources that focus on either isolated cues or high-level social QA, GRASP builds questions from identity-consistent gaze trajectories, deictic gestures, and their joint compositions into social events. Moreover, we propose Social Grounding Reward (SGR), a learning signal that uses these social events to encourage models to reason about the participants involved in each interaction. Experiments show that SGR improves performance on GRASP-Bench while maintaining zero-shot performance on related social video QA benchmarks.

Sung Jin Um, Dongjin Kim, Sangmin Lee et al.

Audio-visual sound source localization task aims to spatially localize sound-making objects within visual scenes by integrating visual and audio cues. However, existing methods struggle with accurately localizing sound-making objects in complex scenes, particularly when visually similar silent objects coexist. This limitation arises primarily from their reliance on simple audio-visual correspondence, which does not capture fine-grained semantic differences between sound-making and silent objects. To address these challenges, we propose a novel sound source localization framework leveraging Multimodal Large Language Models (MLLMs) to generate detailed contextual information that explicitly distinguishes between sound-making foreground objects and silent background objects. To effectively integrate this detailed information, we introduce two novel loss functions: Object-aware Contrastive Alignment (OCA) loss and Object Region Isolation (ORI) loss. Extensive experimental results on MUSIC and VGGSound datasets demonstrate the effectiveness of our approach, significantly outperforming existing methods in both single-source and multi-source localization scenarios. Code and generated detailed contextual information are available at: https://github.com/VisualAIKHU/OA-SSL.

Xu Cao, Pranav Virupaksha, Wenqi Jia et al. · gatech

Previous research in human gesture recognition has largely overlooked multi-person interactions, which are crucial for understanding the social context of naturally occurring gestures. This limitation in existing datasets presents a significant challenge in aligning human gestures with other modalities like language and speech. To address this issue, we introduce SocialGesture, the first large-scale dataset specifically designed for multi-person gesture analysis. SocialGesture features a diverse range of natural scenarios and supports multiple gesture analysis tasks, including video-based recognition and temporal localization, providing a valuable resource for advancing the study of gesture during complex social interactions. Furthermore, we propose a novel visual question answering (VQA) task to benchmark vision language models'(VLMs) performance on social gesture understanding. Our findings highlight several limitations of current gesture recognition models, offering insights into future directions for improvement in this field. SocialGesture is available at huggingface.co/datasets/IrohXu/SocialGesture.

Hongyeob Kim, Inyoung Jung, Dayoon Suh et al.

Audio-Visual Question Answering (AVQA) requires not only question-based multimodal reasoning but also precise temporal grounding to capture subtle dynamics for accurate prediction. However, existing methods mainly use question information implicitly, limiting focus on question-specific details. Furthermore, most studies rely on uniform frame sampling, which can miss key question-relevant frames. Although recent Top-K frame selection methods aim to address this, their discrete nature still overlooks fine-grained temporal details. This paper proposes QA-TIGER, a novel framework that explicitly incorporates question information and models continuous temporal dynamics. Our key idea is to use Gaussian-based modeling to adaptively focus on both consecutive and non-consecutive frames based on the question, while explicitly injecting question information and applying progressive refinement. We leverage a Mixture of Experts (MoE) to flexibly implement multiple Gaussian models, activating temporal experts specifically tailored to the question. Extensive experiments on multiple AVQA benchmarks show that QA-TIGER consistently achieves state-of-the-art performance. Code is available at https://aim-skku.github.io/QA-TIGER/

Sung Jin Um, Dongjin Kim, Sangmin Lee et al.

The goal of video moment retrieval and highlight detection is to identify specific segments and highlights based on a given text query. With the rapid growth of video content and the overlap between these tasks, recent works have addressed both simultaneously. However, they still struggle to fully capture the overall video context, making it challenging to determine which words are most relevant. In this paper, we present a novel Video Context-aware Keyword Attention module that overcomes this limitation by capturing keyword variation within the context of the entire video. To achieve this, we introduce a video context clustering module that provides concise representations of the overall video context, thereby enhancing the understanding of keyword dynamics. Furthermore, we propose a keyword weight detection module with keyword-aware contrastive learning that incorporates keyword information to enhance fine-grained alignment between visual and textual features. Extensive experiments on the QVHighlights, TVSum, and Charades-STA benchmarks demonstrate that our proposed method significantly improves performance in moment retrieval and highlight detection tasks compared to existing approaches. Our code is available at: https://github.com/VisualAIKHU/Keyword-DETR

Joonhyeon Song, Seohwan Yun, Seongho Yoon et al.

This work proposes a novel approach beyond supervised learning for effective pathological image analysis, addressing the challenge of limited robust labeled data. Pathological diagnosis of diseases like cancer has conventionally relied on the evaluation of morphological features by physicians and pathologists. However, recent advancements in compute-aided diagnosis (CAD) systems are gaining significant attention as diagnostic support tools. Although the advancement of deep learning has improved CAD significantly, segmentation models typically require large pixel-level annotated dataset, and such labeling is expensive. Existing studies not based on supervised approaches still struggle with limited generalization, and no practical approach has emerged yet. To address this issue, we present a weakly supervised semantic segmentation (WSSS) model by combining class activation map and Segment Anything Model (SAM)-based pseudo-labeling. For effective pretraining, we adopt the SAM-a foundation model that is pretrained on large datasets and operates in zero-shot configurations using only coarse prompts. The proposed approach transfer enhanced Attention Dropout Layer's knowledge to SAM, thereby generating pseudo-labels. To demonstrate the superiority of the proposed method, experimental studies are conducted on histopathological breast cancer datasets. The proposed method outperformed other WSSS methods across three datasets, demonstrating its efficiency by achieving this with only 12GB of GPU memory during training. Our code is available at : https://github.com/QI-NemoSong/EP-SAM

Kyeonghyun Yoo, Wooyong Jung, Namkyung Yoon et al.

Ultra-wideband (UWB) indoor localization provides centimeter-level accuracy and low latency, but its measurement reliability degrades severely under Non-Line-of-Sight (NLOS) conditions, leading to meter-scale ranging errors and inconsistent uncertainty characteristics. Inertial Measurement Unit (IMU)-based Pedestrian Dead Reckoning (PDR) complements UWB by providing infrastructure-free motion estimation; however, its error accumulates nonlinearly over time due to bias and noise propagation. Fusion methods based on Extended Kalman Filters (EKF) and Particle Filters (PF) can improve average localization accuracy through probabilistic state estimation. However, these approaches typically rely on manually tuned measurement covariances. Such fixed or heuristically tuned parameters are hard to sustain across varying indoor layouts, NLOS ratios, and motion patterns, leading to limited robustness and poor generalization of measurement uncertainty modeling in heterogeneous environments. To address this limitation, this work proposes an adaptive measurement covariance scaling framework in which reliability cues are learned from historical UWB/PDR trajectories. A large teacher model is employed offline to generate temporally consistent next-position predictions from structured UWB/PDR sequences, and this behavior is distilled into a lightweight student model suitable for real-time deployment. The student model continuously regulates EKF measurement covariances based on prediction residuals, enabling environment-aware fusion without manual re-tuning. Experimental results demonstrate that the proposed KD-EKF framework significantly reduces localization error, suppresses error spikes during Line-of-Sight (LOS)/NLOS transitions, and mitigates long-term drift compared to fixed-parameter EKF, thereby improving measurement robustness across diverse indoor environments.

Seongbin Kim, Gyuwan Kim, Seongjin Shin et al.

End-to-end approaches open a new way for more accurate and efficient spoken language understanding (SLU) systems by alleviating the drawbacks of traditional pipeline systems. Previous works exploit textual information for an SLU model via pre-training with automatic speech recognition or fine-tuning with knowledge distillation. To utilize textual information more effectively, this work proposes a two-stage textual knowledge distillation method that matches utterance-level representations and predicted logits of two modalities during pre-training and fine-tuning, sequentially. We use vq-wav2vec BERT as a speech encoder because it captures general and rich features. Furthermore, we improve the performance, especially in a low-resource scenario, with data augmentation methods by randomly masking spans of discrete audio tokens and contextualized hidden representations. Consequently, we push the state-of-the-art on the Fluent Speech Commands, achieving 99.7% test accuracy in the full dataset setting and 99.5% in the 10% subset setting. Throughout the ablation studies, we empirically verify that all used methods are crucial to the final performance, providing the best practice for spoken language understanding. Code is available at https://github.com/clovaai/textual-kd-slu.

Sangmin Lee, Eunpil Park, Angel Canelo et al.

This paper reviews the NTIRE 2025 Efficient Burst HDR and Restoration Challenge, which aims to advance efficient multi-frame high dynamic range (HDR) and restoration techniques. The challenge is based on a novel RAW multi-frame fusion dataset, comprising nine noisy and misaligned RAW frames with various exposure levels per scene. Participants were tasked with developing solutions capable of effectively fusing these frames while adhering to strict efficiency constraints: fewer than 30 million model parameters and a computational budget under 4.0 trillion FLOPs. A total of 217 participants registered, with six teams finally submitting valid solutions. The top-performing approach achieved a PSNR of 43.22 dB, showcasing the potential of novel methods in this domain. This paper provides a comprehensive overview of the challenge, compares the proposed solutions, and serves as a valuable reference for researchers and practitioners in efficient burst HDR and restoration.

Sangmin Lee, Bolin Lai, Fiona Ryan et al. · gatech

Understanding social interactions involving both verbal and non-verbal cues is essential for effectively interpreting social situations. However, most prior works on multimodal social cues focus predominantly on single-person behaviors or rely on holistic visual representations that are not aligned to utterances in multi-party environments. Consequently, they are limited in modeling the intricate dynamics of multi-party interactions. In this paper, we introduce three new challenging tasks to model the fine-grained dynamics between multiple people: speaking target identification, pronoun coreference resolution, and mentioned player prediction. We contribute extensive data annotations to curate these new challenges in social deduction game settings. Furthermore, we propose a novel multimodal baseline that leverages densely aligned language-visual representations by synchronizing visual features with their corresponding utterances. This facilitates concurrently capturing verbal and non-verbal cues pertinent to social reasoning. Experiments demonstrate the effectiveness of the proposed approach with densely aligned multimodal representations in modeling fine-grained social interactions. Project website: https://sangmin-git.github.io/projects/MMSI.

Bolin Lai, Felix Juefei-Xu, Miao Liu et al.

Text-guided image manipulation has experienced notable advancement in recent years. In order to mitigate linguistic ambiguity, few-shot learning with visual examples has been applied for instructions that are underrepresented in the training set, or difficult to describe purely in language. However, learning from visual prompts requires strong reasoning capability, which diffusion models are struggling with. To address this issue, we introduce a novel multi-modal autoregressive model, dubbed $\textbf{InstaManip}$, that can $\textbf{insta}$ntly learn a new image $\textbf{manip}$ulation operation from textual and visual guidance via in-context learning, and apply it to new query images. Specifically, we propose an innovative group self-attention mechanism to break down the in-context learning process into two separate stages -- learning and applying, which simplifies the complex problem into two easier tasks. We also introduce a relation regularization method to further disentangle image transformation features from irrelevant contents in exemplar images. Extensive experiments suggest that our method surpasses previous few-shot image manipulation models by a notable margin ($\geq$19% in human evaluation). We also find our model can be further boosted by increasing the number or diversity of exemplar images.

Jaehoon Ko, Kyusun Cho, Joungbin Lee et al.

Recent methods for audio-driven talking head synthesis often optimize neural radiance fields (NeRF) on a monocular talking portrait video, leveraging its capability to render high-fidelity and 3D-consistent novel-view frames. However, they often struggle to reconstruct complete face geometry due to the absence of comprehensive 3D information in the input monocular videos. In this paper, we introduce a novel audio-driven talking head synthesis framework, called Talk3D, that can faithfully reconstruct its plausible facial geometries by effectively adopting the pre-trained 3D-aware generative prior. Given the personalized 3D generative model, we present a novel audio-guided attention U-Net architecture that predicts the dynamic face variations in the NeRF space driven by audio. Furthermore, our model is further modulated by audio-unrelated conditioning tokens which effectively disentangle variations unrelated to audio features. Compared to existing methods, our method excels in generating realistic facial geometries even under extreme head poses. We also conduct extensive experiments showing our approach surpasses state-of-the-art benchmarks in terms of both quantitative and qualitative evaluations.

Won Jun Kim, Hyungjin Chung, Jaemin Kim et al.

Gradient-based methods are a prototypical family of explainability techniques, especially for image-based models. Nonetheless, they have several shortcomings in that they (1) require white-box access to models, (2) are vulnerable to adversarial attacks, and (3) produce attributions that lie off the image manifold, leading to explanations that are not actually faithful to the model and do not align well with human perception. To overcome these challenges, we introduce Derivative-Free Diffusion Manifold-Constrainted Gradients (FreeMCG), a novel method that serves as an improved basis for explainability of a given neural network than the traditional gradient. Specifically, by leveraging ensemble Kalman filters and diffusion models, we derive a derivative-free approximation of the model's gradient projected onto the data manifold, requiring access only to the model's outputs. We demonstrate the effectiveness of FreeMCG by applying it to both counterfactual generation and feature attribution, which have traditionally been treated as distinct tasks. Through comprehensive evaluation on both tasks, counterfactual explanation and feature attribution, we show that our method yields state-of-the-art results while preserving the essential properties expected of XAI tools.

Sangmin Lee, Woojin Chung, Seyun Um et al.

Code-switching (CS), the alternation between two or more languages within a single speaker's utterances, is common in real-world conversations and poses significant challenges for multilingual speech technology. However, systems capable of handling this phenomenon remain underexplored, primarily due to the scarcity of suitable datasets. To resolve this issue, we propose Universal Code-Mixer (UniCoM), a novel pipeline for generating high-quality, natural CS samples without altering sentence semantics. Our approach utilizes an algorithm we call Substituting WORDs with Synonyms (SWORDS), which generates CS speech by replacing selected words with their translations while considering their parts of speech. Using UniCoM, we construct Code-Switching FLEURS (CS-FLEURS), a multilingual CS corpus designed for automatic speech recognition (ASR) and speech-to-text translation (S2TT). Experimental results show that CS-FLEURS achieves high intelligibility and naturalness, performing comparably to existing datasets on both objective and subjective metrics. We expect our approach to advance CS speech technology and enable more inclusive multilingual systems.

Hyung Kyu Kim, Sangmin Lee, Hak Gu Kim

Speech-driven 3D facial animation aims to synthesize realistic facial motion sequences from given audio, matching the speaker's speaking style. However, previous works often require priors such as class labels of a speaker or additional 3D facial meshes at inference, which makes them fail to reflect the speaking style and limits their practical use. To address these issues, we propose MemoryTalker which enables realistic and accurate 3D facial motion synthesis by reflecting speaking style only with audio input to maximize usability in applications. Our framework consists of two training stages: 1-stage is storing and retrieving general motion (i.e., Memorizing), and 2-stage is to perform the personalized facial motion synthesis (i.e., Animating) with the motion memory stylized by the audio-driven speaking style feature. In this second stage, our model learns about which facial motion types should be emphasized for a particular piece of audio. As a result, our MemoryTalker can generate a reliable personalized facial animation without additional prior information. With quantitative and qualitative evaluations, as well as user study, we show the effectiveness of our model and its performance enhancement for personalized facial animation over state-of-the-art methods.

Sejin Park, Sangmin Lee, Kyong Hwan Jin et al.

Super-resolution (SR) has been a pivotal task in image processing, aimed at enhancing image resolution across various applications. Recently, look-up table (LUT)-based approaches have attracted interest due to their efficiency and performance. However, these methods are typically designed for fixed scale factors, making them unsuitable for arbitrary-scale image SR (ASISR). Existing ASISR techniques often employ implicit neural representations, which come with considerable computational cost and memory demands. To address these limitations, we propose Interpolation Mixing LUT (IM-LUT), a novel framework that operates ASISR by learning to blend multiple interpolation functions to maximize their representational capacity. Specifically, we introduce IM-Net, a network trained to predict mixing weights for interpolation functions based on local image patterns and the target scale factor. To enhance efficiency of interpolation-based methods, IM-Net is transformed into IM-LUT, where LUTs are employed to replace computationally expensive operations, enabling lightweight and fast inference on CPUs while preserving reconstruction quality. Experimental results on several benchmark datasets demonstrate that IM-LUT consistently achieves a superior balance between image quality and efficiency compared to existing methods, highlighting its potential as a promising solution for resource-constrained applications.

Sangmin Lee, Abbas Mammadov, Jong Chul Ye

Current theoretical and empirical research in neural networks suggests that complex datasets require large network architectures for thorough classification, yet the precise nature of this relationship remains unclear. This paper tackles this issue by defining upper and lower bounds for neural network widths, which are informed by the polytope structure of the dataset in question. We also delve into the application of these principles to simplicial complexes and specific manifold shapes, explaining how the requirement for network width varies in accordance with the geometric complexity of the dataset. Moreover, we develop an algorithm to investigate a converse situation where the polytope structure of a dataset can be inferred from its corresponding trained neural networks. Through our algorithm, it is established that popular datasets such as MNIST, Fashion-MNIST, and CIFAR10 can be efficiently encapsulated using no more than two polytopes with a small number of faces.

Gyumin Shim, Sangmin Lee, Jaegul Choo

In this paper, we introduce GaussianMotion, a novel human rendering model that generates fully animatable scenes aligned with textual descriptions using Gaussian Splatting. Although existing methods achieve reasonable text-to-3D generation of human bodies using various 3D representations, they often face limitations in fidelity and efficiency, or primarily focus on static models with limited pose control. In contrast, our method generates fully animatable 3D avatars by combining deformable 3D Gaussian Splatting with text-to-3D score distillation, achieving high fidelity and efficient rendering for arbitrary poses. By densely generating diverse random poses during optimization, our deformable 3D human model learns to capture a wide range of natural motions distilled from a pose-conditioned diffusion model in an end-to-end manner. Furthermore, we propose Adaptive Score Distillation that effectively balances realistic detail and smoothness to achieve optimal 3D results. Experimental results demonstrate that our approach outperforms existing baselines by producing high-quality textures in both static and animated results, and by generating diverse 3D human models from various textual inputs.

Woongjib Choi, Sangmin Lee, Hyungseob Lim et al.

In this paper, we present a vocoder-free framework for audio super-resolution that employs a flow matching generative model to capture the conditional distribution of complex-valued spectral coefficients. Unlike conventional two-stage diffusion-based approaches that predict a mel-spectrogram and then rely on a pre-trained neural vocoder to synthesize waveforms, our method directly reconstructs waveforms via the inverse Short-Time Fourier Transform (iSTFT), thereby eliminating the dependence on a separate vocoder. This design not only simplifies end-to-end optimization but also overcomes a critical bottleneck of two-stage pipelines, where the final audio quality is fundamentally constrained by vocoder performance. Experiments show that our model consistently produces high-fidelity 48 kHz audio across diverse upsampling factors, achieving state-of-the-art performance on both speech and general audio datasets.

Sangmin Lee, Woongjib Choi, Jihyun Kim et al.

In this paper, we present a neural spoken language diarization model that supports an unconstrained span of languages within a single framework. Our approach integrates a learnable query-based architecture grounded in multilingual awareness, with large-scale pretraining on simulated code-switching data. By jointly leveraging these two components, our method overcomes the limitations of conventional approaches in data scarcity and architecture optimization, and generalizes effectively to real-world multilingual settings across diverse environments. Experimental results demonstrate that our approach achieves state-of-the-art performance on several language diarization benchmarks, with a relative performance improvement of 23% to 52% over previous methods. We believe that this work not only advances research in language diarization but also establishes a foundational framework for code-switching speech technologies.

Gyeongje Cho, Yeonkyoun So, Chanwoo Park et al.

This paper introduces Thunder-Tok, a new Korean tokenizer designed to reduce token fertility without compromising model performance. Our approach uses a rule-based pre-tokenization method that aligns with the linguistic structure of the Korean language. We also create a seed vocabulary containing tokens that resemble linguistic units and employ a branching entropy-based selection algorithm. These techniques increase the average token length, thus lowering fertility while preserving linguistic information. Experimental results indicate that Thunder-Tok reduces fertility by approximately 10% (i.e., reduces the number of tokens by 10%, improving the inference speed by 10%) compared to BPE without compromising performance across various downstream tasks. These findings demonstrate that our linguistically informed approach is effective and practical for designing efficient tokenizers for language models.

Bolin Lai, Sangmin Lee, Xu Cao et al.

Text-image-to-video (TI2V) generation is a critical problem for controllable video generation using both semantic and visual conditions. Most existing methods typically add visual conditions to text-to-video (T2V) foundation models by finetuning, which is costly in resources and only limited to a few predefined conditioning settings. To tackle this issue, we introduce a unified formulation for TI2V generation with flexible visual conditioning. Furthermore, we propose an innovative training-free approach, dubbed FlexTI2V, that can condition T2V foundation models on an arbitrary amount of images at arbitrary positions. Specifically, we firstly invert the condition images to noisy representation in a latent space. Then, in the denoising process of T2V models, our method uses a novel random patch swapping strategy to incorporate visual features into video representations through local image patches. To balance creativity and fidelity, we use a dynamic control mechanism to adjust the strength of visual conditioning to each video frame. Extensive experiments validate that our method surpasses previous training-free image conditioning methods by a notable margin. We also show more insights of our method by detailed ablation study and analysis.

Sangmin Lee, Woo-Jin Chung, Hong-Goo Kang

Building a universal multilingual automatic speech recognition (ASR) model that performs equitably across languages has long been a challenge due to its inherent difficulties. To address this task we introduce a Language-Agnostic Multilingual ASR pipeline through orthography Unification and language-specific Transliteration (LAMA-UT). LAMA-UT operates without any language-specific modules while matching the performance of state-of-the-art models trained on a minimal amount of data. Our pipeline consists of two key steps. First, we utilize a universal transcription generator to unify orthographic features into Romanized form and capture common phonetic characteristics across diverse languages. Second, we utilize a universal converter to transform these universal transcriptions into language-specific ones. In experiments, we demonstrate the effectiveness of our proposed method leveraging universal transcriptions for massively multilingual ASR. Our pipeline achieves a relative error reduction rate of 45% when compared to Whisper and performs comparably to MMS, despite being trained on only 0.1% of Whisper's training data. Furthermore, our pipeline does not rely on any language-specific modules. However, it performs on par with zero-shot ASR approaches which utilize additional language-specific lexicons and language models. We expect this framework to serve as a cornerstone for flexible multilingual ASR systems that are generalizable even to unseen languages.

Sangmin Lee, Jong Chul Ye

This paper investigates the relationship between the universal approximation property of deep neural networks and topological characteristics of datasets. Our primary contribution is to introduce data topology-dependent upper bounds on the network width. Specifically, we first show that a three-layer neural network, applying a ReLU activation function and max pooling, can be designed to approximate an indicator function over a compact set, one that is encompassed by a tight convex polytope. This is then extended to a simplicial complex, deriving width upper bounds based on its topological structure. Further, we calculate upper bounds in relation to the Betti numbers of select topological spaces. Finally, we prove the universal approximation property of three-layer ReLU networks using our topological approach. We also verify that gradient descent converges to the network structure proposed in our study.

Younghun Cho, Giseop Kim, Sangmin Lee et al.

Using publicly accessible maps, we propose a novel vehicle localization method that can be applied without using prior light detection and ranging (LiDAR) maps. Our method generates OSM descriptors by calculating the distances to buildings from a location in OpenStreetMap at a regular angle, and LiDAR descriptors by calculating the shortest distances to building points from the current location at a regular angle. Comparing the OSM descriptors and LiDAR descriptors yields a highly accurate vehicle localization result. Compared to methods that use prior LiDAR maps, our method presents two main advantages: (1) vehicle localization is not limited to only places with previously acquired LiDAR maps, and (2) our method is comparable to LiDAR map-based methods, and especially outperforms the other methods with respect to the top one candidate at KITTI dataset sequence 00.

Sangmin Lee, Byeongsu Sim, Jong Chul Ye

Understanding implicit bias of gradient descent for generalization capability of ReLU networks has been an important research topic in machine learning research. Unfortunately, even for a single ReLU neuron trained with the square loss, it was recently shown impossible to characterize the implicit regularization in terms of a norm of model parameters (Vardi & Shamir, 2021). In order to close the gap toward understanding intriguing generalization behavior of ReLU networks, here we examine the gradient flow dynamics in the parameter space when training single-neuron ReLU networks. Specifically, we discover an implicit bias in terms of support vectors, which plays a key role in why and how ReLU networks generalize well. Moreover, we analyze gradient flows with respect to the magnitude of the norm of initialization, and show that the norm of the learned weight strictly increases through the gradient flow. Lastly, we prove the global convergence of single ReLU neuron for $d = 2$ case.

Hak Gu Kim, Minho Park, Sangmin Lee et al.

Depth adjustment aims to enhance the visual experience of stereoscopic 3D (S3D) images, which accompanied with improving visual comfort and depth perception. For a human expert, the depth adjustment procedure is a sequence of iterative decision making. The human expert iteratively adjusts the depth until he is satisfied with the both levels of visual comfort and the perceived depth. In this work, we present a novel deep reinforcement learning (DRL)-based approach for depth adjustment named VCA-RL (Visual Comfort Aware Reinforcement Learning) to explicitly model human sequential decision making in depth editing operations. We formulate the depth adjustment process as a Markov decision process where actions are defined as camera movement operations to control the distance between the left and right cameras. Our agent is trained based on the guidance of an objective visual comfort assessment metric to learn the optimal sequence of camera movement actions in terms of perceptual aspects in stereoscopic viewing. With extensive experiments and user studies, we show the effectiveness of our VCA-RL model on three different S3D databases.

Hak Gu Kim, Sangmin Lee, Seongyeop Kim et al.

We address the black-box issue of VR sickness assessment (VRSA) by evaluating the level of physical symptoms of VR sickness. For the VR contents inducing the similar VR sickness level, the physical symptoms can vary depending on the characteristics of the contents. Most of existing VRSA methods focused on assessing the overall VR sickness score. To make better understanding of VR sickness, it is required to predict and provide the level of major symptoms of VR sickness rather than overall degree of VR sickness. In this paper, we predict the degrees of main physical symptoms affecting the overall degree of VR sickness, which are disorientation, nausea, and oculomotor. In addition, we introduce a new large-scale dataset for VRSA including 360 videos with various frame rates, physiological signals, and subjective scores. On VRSA benchmark and our newly collected dataset, our approach shows a potential to not only achieve the highest correlation with subjective scores, but also to better understand which symptoms are the main causes of VR sickness.

Sangmin Lee, Hak Gu Kim, Dae Hwi Choi et al.

Our work addresses long-term motion context issues for predicting future frames. To predict the future precisely, it is required to capture which long-term motion context (e.g., walking or running) the input motion (e.g., leg movement) belongs to. The bottlenecks arising when dealing with the long-term motion context are: (i) how to predict the long-term motion context naturally matching input sequences with limited dynamics, (ii) how to predict the long-term motion context with high-dimensionality (e.g., complex motion). To address the issues, we propose novel motion context-aware video prediction. To solve the bottleneck (i), we introduce a long-term motion context memory (LMC-Memory) with memory alignment learning. The proposed memory alignment learning enables to store long-term motion contexts into the memory and to match them with sequences including limited dynamics. As a result, the long-term context can be recalled from the limited input sequence. In addition, to resolve the bottleneck (ii), we propose memory query decomposition to store local motion context (i.e., low-dimensional dynamics) and recall the suitable local context for each local part of the input individually. It enables to boost the alignment effects of the memory. Experimental results show that the proposed method outperforms other sophisticated RNN-based methods, especially in long-term condition. Further, we validate the effectiveness of the proposed network designs by conducting ablation studies and memory feature analysis. The source code of this work is available.

Sam Sattarzadeh, Mahesh Sudhakar, Anthony Lem et al.

As an emerging field in Machine Learning, Explainable AI (XAI) has been offering remarkable performance in interpreting the decisions made by Convolutional Neural Networks (CNNs). To achieve visual explanations for CNNs, methods based on class activation mapping and randomized input sampling have gained great popularity. However, the attribution methods based on these techniques provide lower resolution and blurry explanation maps that limit their explanation power. To circumvent this issue, visualization based on various layers is sought. In this work, we collect visualization maps from multiple layers of the model based on an attribution-based input sampling technique and aggregate them to reach a fine-grained and complete explanation. We also propose a layer selection strategy that applies to the whole family of CNN-based models, based on which our extraction framework is applied to visualize the last layers of each convolutional block of the model. Moreover, we perform an empirical analysis of the efficacy of derived lower-level information to enhance the represented attributions. Comprehensive experiments conducted on shallow and deep models trained on natural and industrial datasets, using both ground-truth and model-truth based evaluation metrics validate our proposed algorithm by meeting or outperforming the state-of-the-art methods in terms of explanation ability and visual quality, demonstrating that our method shows stability regardless of the size of objects or instances to be explained.

Joanna Hong, Jung Uk Kim, Sangmin Lee et al.

Recent advances in facial expression synthesis have shown promising results using diverse expression representations including facial action units. Facial action units for an elaborate facial expression synthesis need to be intuitively represented for human comprehension, not a numeric categorization of facial action units. To address this issue, we utilize human-friendly approach: use of natural language where language helps human grasp conceptual contexts. In this paper, therefore, we propose a new facial expression synthesis model from language-based facial expression description. Our method can synthesize the facial image with detailed expressions. In addition, effectively embedding language features on facial features, our method can control individual word to handle each part of facial movement. Extensive qualitative and quantitative evaluations were conducted to verify the effectiveness of the natural language.

Sangmin Lee, Hak Gu Kim, Yong Man Ro

In this paper, we propose a novel abnormal event detection method with spatio-temporal adversarial networks (STAN). We devise a spatio-temporal generator which synthesizes an inter-frame by considering spatio-temporal characteristics with bidirectional ConvLSTM. A proposed spatio-temporal discriminator determines whether an input sequence is real-normal or not with 3D convolutional layers. These two networks are trained in an adversarial way to effectively encode spatio-temporal features of normal patterns. After the learning, the generator and the discriminator can be independently used as detectors, and deviations from the learned normal patterns are detected as abnormalities. Experimental results show that the proposed method achieved competitive performance compared to the state-of-the-art methods. Further, for the interpretation, we visualize the location of abnormal events detected by the proposed networks using a generator loss and discriminator gradients.

Wissam J. Baddar, Geonmo Gu, Sangmin Lee et al.

In this paper, we propose Dynamics Transfer GAN; a new method for generating video sequences based on generative adversarial learning. The spatial constructs of a generated video sequence are acquired from the target image. The dynamics of the generated video sequence are imported from a source video sequence, with arbitrary motion, and imposed onto the target image. To preserve the spatial construct of the target image, the appearance of the source video sequence is suppressed and only the dynamics are obtained before being imposed onto the target image. That is achieved using the proposed appearance suppressed dynamics feature. Moreover, the spatial and temporal consistencies of the generated video sequence are verified via two discriminator networks. One discriminator validates the fidelity of the generated frames appearance, while the other validates the dynamic consistency of the generated video sequence. Experiments have been conducted to verify the quality of the video sequences generated by the proposed method. The results verified that Dynamics Transfer GAN successfully transferred arbitrary dynamics of the source video sequence onto a target image when generating the output video sequence. The experimental results also showed that Dynamics Transfer GAN maintained the spatial constructs (appearance) of the target image while generating spatially and temporally consistent video sequences.