Hyesong Choi

Semi Lee, Hyejin Go, Hyesong Choi

The quadratic cost of self-attention in Vision Transformers (ViTs) constitutes a fundamental bottleneck for practical deployment, motivating a vibrant line of research on token reduction. Among existing approaches, token merging (ToMe) has emerged as an elegant training-free solution; yet its design rests on an unspoken premise of token equality, which contravenes the well-documented non-uniformity of self-attention and leads to information loss in high-salience tokens under aggressive compression. We address this limitation with AdaMerge, a token-merging framework based on two complementary mechanisms. First, salience-weighted similarity leverages column-wise feature-affinity centrality as a token-importance proxy and incorporates the resulting salience scores into the bipartite matching score, ensuring that pivotal tokens contribute more strongly to the merged representation. Second, adaptive merging intensity uses pre-computed layer-wise similarity statistics to dynamically modulate the per-layer reduction count in accordance with input-specific redundancy. On ImageNet-1k with ViT-B/16, AdaMerge consistently outperforms ToMe, PiToMe, and DSM across all FLOPs-matched regimes. The accuracy gap widens monotonically with compression: at the 13.4G FLOPs operating point, AdaMerge sustains a Top-1 degradation of only -1.06%, compared to -1.45% for PiToMe and -4.62% for DSM. To our knowledge, AdaMerge is the first to combine salience-weighted similarity and adaptive per-layer reduction into a single training-free token merging framework, advancing the accuracy-FLOPs Pareto frontier of ViT acceleration.

Sunkyung Kim, Hyesong Choi, Dongbo Min

In multi-task learning (MTL) for visual scene understanding, it is crucial to transfer useful information between multiple tasks with minimal interferences. In this paper, we propose a novel architecture that effectively transfers informative features by applying the attention mechanism to the multi-scale features of the tasks. Since applying the attention module directly to all possible features in terms of scale and task requires a high complexity, we propose to apply the attention module sequentially for the task and scale. The cross-task attention module (CTAM) is first applied to facilitate the exchange of relevant information between the multiple task features of the same scale. The cross-scale attention module (CSAM) then aggregates useful information from feature maps at different resolutions in the same task. Also, we attempt to capture long range dependencies through the self-attention module in the feature extraction network. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the NYUD-v2 and PASCAL-Context dataset.

Kahyeon Nam, Hyesong Choi

Deploying Vision-Language Models on resource-constrained hardware typically requires INT8 quantization, but in joint-embedding architectures such as CLIP this introduces a failure mode distinct from quantized CNN classifiers: activation noise accumulated across transformer blocks perturbs the direction of the multimodal embedding, eroding the cosine alignment on which zero-shot retrieval depends. We characterize this as Quantization-Induced Representation Collapse (QIRC) and quantify it on INT8 CLIP ViT-B/32, where the layer-wise noise-to-signal ratio grows from below 10% in shallow blocks to 52% at Layer 11. We propose LRA-EE (Layer-wise Representation-Aware Early Exit), which bypasses noise-saturated deep layers via Spatio-Semantic Aggregation (replacing the immature shallow [CLS] with a global patch-token average), a learned multi-feature gate (confidence, top-2 margin, spatial-activation variance), and Layer-adaptive Confidence Thresholding calibrated to each layer's Information-to-Noise Ratio. On ImageNet-1K zero-shot classification, LRA-EE reduces FLOPs by 13.4% and improves Top-1 accuracy by +2.44%p (58.72% -> 61.16%) over the INT8 baseline. A four-quadrant decomposition isolates the Rescue Effect: 9.5% of samples are correctly classified at shallow exits but lost to noise at full depth, against only 7.1% suffering the inverse.

Kim Jihyeon, Sohee Kim, Soosan Lee et al.

Recent generative models have largely closed the gap on low-level artifacts - pixel fingerprints, frequency anomalies, upsampling traces - particularly in person-centric and partial-edit settings where the manipulated region is small and surrounded by photometrically authentic content. We introduce Social Gaze Consistency, a high-level semantic cue defined as the mutual coherence of gaze direction, head-eye alignment, and pupil placement between interacting individuals, and show that it constitutes a previously underutilized detection axis orthogonal to existing low-level paradigms. We instantiate this insight through three coupled mechanisms: (i) a controlled diagnostic dataset with region-specific perturbations of gaze-consistent imagery, where strict pair-level grouping forecloses generator-fingerprint memorization as an optimization-time shortcut rather than relying on augmentation; (ii) Block-Compositional Caption Supervision, which holds a single 5-block reasoning skeleton invariant across 1,250 macro-combined captions, decoupling reasoning consistency from surface diversity; (iii) Cross-architecture validation showing the same supervision improves a vision-language backbone (FakeVLM) by +3.7 pp on the COCOAI Interaction subset (balanced accuracy 67.8 -> 71.5) and +1.3 pp on the COCOAI Person subset (83.0 -> 84.3), with consistent gains on a vision-only backbone (Effort), evidencing a backbone-agnostic cue. Real- and fake-class recalls rise simultaneously, ruling out a "predict-all-fake" artifact. A four-step mechanistic account - paired-edit shortcut blocking, hard-to-easy difficulty transfer, CLIP prior preservation, and diffusion-family shared spectral weakness in periocular structure - explains why training on a single inpainter (FLUX.1-Fill) transfers to multi-generator suites. We will release the code upon acceptance to facilitate reproducibility.

Hyejin Go, Semi Lee, Hyesong Choi

CLIP-style contrastive pretraining typically curates web-scale image-text pairs using sample-level filtering signals, often based on pair-level alignment. We show that this signal saturates: once coarse mismatches are removed, stricter global filtering no longer tracks the compositional supervision provided by the retained captions. The reason is structural - a global score conflates whether a pair is broadly plausible with whether the individual object, attribute, and relation phrases inside the caption materially support the image-text match. The latter is what compositional generalization demands, yet pair-level filters are blind to it. We address this with Counterfactual Phrase Intervention (CPI), a phrase-level curation framework that converts controlled nonce-token substitutions into image-conditioned phrase-sensitivity scores. CPI uses global alignment only for coarse mismatch removal, then ranks the surviving pool by whether caption phrases measurably affect the image-text score under controlled substitution. We frame CPI as a first-order phrase-sensitivity signal rather than a grounding or identification result, and evaluate it at CC3M scale. Ranking by this signal yields a 50%-data subset that improves VL-CheckList-VG Relation by +1.91 over the full-data baseline and +1.00 over alignment-only filtering at matched budget, while improving SugarCrepe overall and preserving general transfer. CPI is loss-orthogonal: applied unchanged to NegCLIP, it further improves VL-CheckList-VG Relation by +3.84, with additional CE-CLIP gains in the main text.

Jiyeong Kim, Yerim So, Hyesong Choi et al.

Unified Multimodal Models (UMMs) have emerged as a promising paradigm that integrates multimodal understanding and generation within a unified modeling framework. However, current generative training paradigms suffer from inherent limitations. We present Semantically-Grounded Supervision (SeGroS), a fine-tuning framework designed to resolve the granularity mismatch and supervisory redundancy in UMMs. At its core, we propose a novel visual grounding map to construct two complementary supervision signals. First, we formulate semantic Visual Hints to compensate for the sparsity of text prompts. Second, we generate a semantically-grounded Corrupted Input to explicitly enhance the supervision of masking-based UMMs by restricting the reconstruction loss to core text-aligned regions. Extensive evaluations on GenEval, DPGBench, and CompBench demonstrate that SeGroS significantly improves generation fidelity and cross-modal alignment across various UMM architectures.

Hayeon Jo, Hyesong Choi, Minhee Cho et al.

Transfer learning based on full fine-tuning (FFT) of the pre-trained encoder and task-specific decoder becomes increasingly complex as deep models grow exponentially. Parameter efficient fine-tuning (PEFT) approaches using adapters consisting of small learnable layers have emerged as an alternative to FFT, achieving comparable performance while maintaining high training efficiency. However, the inflexibility of the adapter with respect to input instances limits its capability of learning task-specific information in diverse downstream tasks. In this paper, we propose a novel PEFT approach, input-Conditioned transFormer, termed iConFormer, that leverages a dynamic adapter conditioned on the input instances. To secure flexible learning ability on input instances in various downstream tasks, we introduce an input-Conditioned Network (iCoN) in the dynamic adapter that enables instance-level feature transformation. To be specific, iCoN generates channel-wise convolutional kernels for each feature and transform it using adaptive convolution process to effectively capture task-specific and fine-grained details tailor to downstream tasks. Experimental results demonstrate that by tuning just 1.6% to 2.8% of the Transformer backbone parameters, iConFormer achieves performance comparable to FFT in monocular depth estimation and semantic segmentation, while outperforming it in image classification and instance segmentation. Also, the proposed method consistently outperforms recent PEFT methods for all the tasks mentioned above.

Jihye Ahn, Hyesong Choi, Soomin Kim et al.

In stereo matching, CNNs have traditionally served as the predominant architectures. Although Transformer-based stereo models have been studied recently, their performance still lags behind CNN-based stereo models due to the inherent data scarcity issue in the stereo matching task. In this paper, we propose Masked Image Modeling Distilled Stereo matching model, termed MaDis-Stereo, that enhances locality inductive bias by leveraging Masked Image Modeling (MIM) in training Transformer-based stereo model. Given randomly masked stereo images as inputs, our method attempts to conduct both image reconstruction and depth prediction tasks. While this strategy is beneficial to resolving the data scarcity issue, the dual challenge of reconstructing masked tokens and subsequently performing stereo matching poses significant challenges, particularly in terms of training stability. To address this, we propose to use an auxiliary network (teacher), updated via Exponential Moving Average (EMA), along with the original stereo model (student), where teacher predictions serve as pseudo supervisory signals to effectively distill knowledge into the student model. State-of-the-arts performance is achieved with the proposed method on several stereo matching such as ETH3D and KITTI 2015. Additionally, to demonstrate that our model effectively leverages locality inductive bias, we provide the attention distance measurement.

Sumin Son, Hyesong Choi, Dongbo Min

Masked Image Modeling (MIM) techniques have redefined the landscape of computer vision, enabling pre-trained models to achieve exceptional performance across a broad spectrum of tasks. Despite their success, the full potential of MIM-based methods in dense prediction tasks, particularly in depth estimation, remains untapped. Existing MIM approaches primarily rely on single-image inputs, which makes it challenging to capture the crucial structured information, leading to suboptimal performance in tasks requiring fine-grained feature representation. To address these limitations, we propose SG-MIM, a novel Structured knowledge Guided Masked Image Modeling framework designed to enhance dense prediction tasks by utilizing structured knowledge alongside images. SG-MIM employs a lightweight relational guidance framework, allowing it to guide structured knowledge individually at the feature level rather than naively combining at the pixel level within the same architecture, as is common in traditional multi-modal pre-training methods. This approach enables the model to efficiently capture essential information while minimizing discrepancies between pre-training and downstream tasks. Furthermore, SG-MIM employs a selective masking strategy to incorporate structured knowledge, maximizing the synergy between general representation learning and structured knowledge-specific learning. Our method requires no additional annotations, making it a versatile and efficient solution for a wide range of applications. Our evaluations on the KITTI, NYU-v2, and ADE20k datasets demonstrate SG-MIM's superiority in monocular depth estimation and semantic segmentation.

Minhee Cho, Hyesong Choi, Hayeon Jo et al.

Unsupervised Domain Adaptation (UDA) endeavors to bridge the gap between a model trained on a labeled source domain and its deployment in an unlabeled target domain. However, current high-performance models demand significant resources, making deployment costs prohibitive and highlighting the need for compact, yet effective models. For UDA of lightweight models, Knowledge Distillation (KD) leveraging a Teacher-Student framework could be a common approach, but we found that domain shift in UDA leads to a significant increase in non-salient parameters in the teacher model, degrading model's generalization ability and transferring misleading information to the student model. Interestingly, we observed that this phenomenon occurs considerably less in the student model. Driven by this insight, we introduce Collaborative Learning for UDA (CLDA), a method that updates the teacher's non-salient parameters using the student model and at the same time utilizes the updated teacher model to improve UDA performance of the student model. Experiments show consistent performance improvements for both student and teacher models. For example, in semantic segmentation, CLDA achieves an improvement of +0.7% mIoU for the teacher model and +1.4% mIoU for the student model compared to the baseline model in the GTA-to-Cityscapes datasets. In the Synthia-to-Cityscapes dataset, it achieves an improvement of +0.8% mIoU and +2.0% mIoU for the teacher and student models, respectively.

Soomin Kim, Hyesong Choi, Jihye Ahn et al.

Unlike other vision tasks where Transformer-based approaches are becoming increasingly common, stereo depth estimation is still dominated by convolution-based approaches. This is mainly due to the limited availability of real-world ground truth for stereo matching, which is a limiting factor in improving the performance of Transformer-based stereo approaches. In this paper, we propose UniTT-Stereo, a method to maximize the potential of Transformer-based stereo architectures by unifying self-supervised learning used for pre-training with stereo matching framework based on supervised learning. To be specific, we explore the effectiveness of reconstructing features of masked portions in an input image and at the same time predicting corresponding points in another image from the perspective of locality inductive bias, which is crucial in training models with limited training data. Moreover, to address these challenging tasks of reconstruction-and-prediction, we present a new strategy to vary a masking ratio when training the stereo model with stereo-tailored losses. State-of-the-art performance of UniTT-Stereo is validated on various benchmarks such as ETH3D, KITTI 2012, and KITTI 2015 datasets. Lastly, to investigate the advantages of the proposed approach, we provide a frequency analysis of feature maps and the analysis of locality inductive bias based on attention maps.

Jueun Ko, Hyewon Park, Hyesong Choi et al.

Stereo Depth Estimation in real-world environments poses significant challenges due to dynamic domain shifts, sparse or unreliable supervision, and the high cost of acquiring dense ground-truth labels. While recent Test-Time Adaptation (TTA) methods offer promising solutions, most rely on static target domain assumptions and input-invariant adaptation strategies, limiting their effectiveness under continual shifts. In this paper, we propose RobIA, a novel Robust, Instance-Aware framework for Continual Test-Time Adaptation (CTTA) in stereo depth estimation. RobIA integrates two key components: (1) Attend-and-Excite Mixture-of-Experts (AttEx-MoE), a parameter-efficient module that dynamically routes input to frozen experts via lightweight self-attention mechanism tailored to epipolar geometry, and (2) Robust AdaptBN Teacher, a PEFT-based teacher model that provides dense pseudo-supervision by complementing sparse handcrafted labels. This strategy enables input-specific flexibility, broad supervision coverage, improving generalization under domain shift. Extensive experiments demonstrate that RobIA achieves superior adaptation performance across dynamic target domains while maintaining computational efficiency.

Hyesong Choi, Hunsang Lee, Seyoung Joung et al.

Driven by the success of Masked Language Modeling (MLM), the realm of self-supervised learning for computer vision has been invigorated by the central role of Masked Image Modeling (MIM) in driving recent breakthroughs. Notwithstanding the achievements of MIM across various downstream tasks, its overall efficiency is occasionally hampered by the lengthy duration of the pre-training phase. This paper presents a perspective that the optimization of masked tokens as a means of addressing the prevailing issue. Initially, we delve into an exploration of the inherent properties that a masked token ought to possess. Within the properties, we principally dedicated to articulating and emphasizing the `data singularity' attribute inherent in masked tokens. Through a comprehensive analysis of the heterogeneity between masked tokens and visible tokens within pre-trained models, we propose a novel approach termed masked token optimization (MTO), specifically designed to improve model efficiency through weight recalibration and the enhancement of the key property of masked tokens. The proposed method serves as an adaptable solution that seamlessly integrates into any MIM approach that leverages masked tokens. As a result, MTO achieves a considerable improvement in pre-training efficiency, resulting in an approximately 50% reduction in pre-training epochs required to attain converged performance of the recent approaches.

Hyesong Choi, Hyejin Park, Kwang Moo Yi et al.

In this paper, we introduce Saliency-Based Adaptive Masking (SBAM), a novel and cost-effective approach that significantly enhances the pre-training performance of Masked Image Modeling (MIM) approaches by prioritizing token salience. Our method provides robustness against variations in masking ratios, effectively mitigating the performance instability issues common in existing methods. This relaxes the sensitivity of MIM-based pre-training to masking ratios, which in turn allows us to propose an adaptive strategy for `tailored' masking ratios for each data sample, which no existing method can provide. Toward this goal, we propose an Adaptive Masking Ratio (AMR) strategy that dynamically adjusts the proportion of masking for the unique content of each image based on token salience. We show that our method significantly improves over the state-of-the-art in mask-based pre-training on the ImageNet-1K dataset.

Seungmin Baek, Soyul Lee, Hayeon Jo et al.

Transfer learning paradigm has driven substantial advancements in various vision tasks. However, as state-of-the-art models continue to grow, classical full fine-tuning often becomes computationally impractical, particularly in multi-task learning (MTL) setup where training complexity increases proportional to the number of tasks. Consequently, recent studies have explored Parameter-Efficient Fine-Tuning (PEFT) for MTL architectures. Despite some progress, these approaches still exhibit limitations in capturing fine-grained, task-specific features that are crucial to MTL. In this paper, we introduce Task-Adaptive Dynamic transFormer, termed TADFormer, a novel PEFT framework that performs task-aware feature adaptation in the fine-grained manner by dynamically considering task-specific input contexts. TADFormer proposes the parameter-efficient prompting for task adaptation and the Dynamic Task Filter (DTF) to capture task information conditioned on input contexts. Experiments on the PASCAL-Context benchmark demonstrate that the proposed method achieves higher accuracy in dense scene understanding tasks, while reducing the number of trainable parameters by up to 8.4 times when compared to full fine-tuning of MTL models. TADFormer also demonstrates superior parameter efficiency and accuracy compared to recent PEFT methods.

Hyesong Choi, Daeun Kim, Sungmin Cha et al.

In this work, we dive deep into the impact of additive noise in pre-training deep networks. While various methods have attempted to use additive noise inspired by the success of latent denoising diffusion models, when used in combination with masked image modeling, their gains have been marginal when it comes to recognition tasks. We thus investigate why this would be the case, in an attempt to find effective ways to combine the two ideas. Specifically, we find three critical conditions: corruption and restoration must be applied within the encoder, noise must be introduced in the feature space, and an explicit disentanglement between noised and masked tokens is necessary. By implementing these findings, we demonstrate improved pre-training performance for a wide range of recognition tasks, including those that require fine-grained, high-frequency information to solve.

Hyesong Choi, Hunsang Lee, Sunkyung Kim et al.

Self-supervised monocular depth estimation has become an appealing solution to the lack of ground truth labels, but its reconstruction loss often produces over-smoothed results across object boundaries and is incapable of handling occlusion explicitly. In this paper, we propose a new approach to leverage pseudo ground truth depth maps of stereo images generated from self-supervised stereo matching methods. The confidence map of the pseudo ground truth depth map is estimated to mitigate performance degeneration by inaccurate pseudo depth maps. To cope with the prediction error of the confidence map itself, we also leverage the threshold network that learns the threshold dynamically conditioned on the pseudo depth maps. The pseudo depth labels filtered out by the thresholded confidence map are used to supervise the monocular depth network. Furthermore, we propose the probabilistic framework that refines the monocular depth map with the help of its uncertainty map through the pixel-adaptive convolution (PAC) layer. Experimental results demonstrate superior performance to state-of-the-art monocular depth estimation methods. Lastly, we exhibit that the proposed threshold learning can also be used to improve the performance of existing confidence estimation approaches.