Junmo Kim

Janghyeon Lee, Jongsuk Kim, Hyounguk Shon et al. · nvidia, utoronto

Pre-training vision-language models with contrastive objectives has shown promising results that are both scalable to large uncurated datasets and transferable to many downstream applications. Some following works have targeted to improve data efficiency by adding self-supervision terms, but inter-domain (image-text) contrastive loss and intra-domain (image-image) contrastive loss are defined on individual spaces in those works, so many feasible combinations of supervision are overlooked. To overcome this issue, we propose UniCLIP, a Unified framework for Contrastive Language-Image Pre-training. UniCLIP integrates the contrastive loss of both inter-domain pairs and intra-domain pairs into a single universal space. The discrepancies that occur when integrating contrastive loss between different domains are resolved by the three key components of UniCLIP: (1) augmentation-aware feature embedding, (2) MP-NCE loss, and (3) domain dependent similarity measure. UniCLIP outperforms previous vision-language pre-training methods on various single- and multi-modality downstream tasks. In our experiments, we show that each component that comprises UniCLIP contributes well to the final performance.

Hanbyel Cho, Junmo Kim

This work focuses on the problem of reconstructing a 3D human body mesh from a given 2D image. Despite the inherent ambiguity of the task of human mesh recovery, most existing works have adopted a method of regressing a single output. In contrast, we propose a generative approach framework, called "Diffusion-based Human Mesh Recovery (Diff-HMR)" that takes advantage of the denoising diffusion process to account for multiple plausible outcomes. During the training phase, the SMPL parameters are diffused from ground-truth parameters to random distribution, and Diff-HMR learns the reverse process of this diffusion. In the inference phase, the model progressively refines the given random SMPL parameters into the corresponding parameters that align with the input image. Diff-HMR, being a generative approach, is capable of generating diverse results for the same input image as the input noise varies. We conduct validation experiments, and the results demonstrate that the proposed framework effectively models the inherent ambiguity of the task of human mesh recovery in a probabilistic manner. The code is available at https://github.com/hanbyel0105/Diff-HMR

Taehoon Kim, Pyunghwan Ahn, Sangyun Kim et al. · nvidia, utoronto

In this report, we introduce NICE (New frontiers for zero-shot Image Captioning Evaluation) project and share the results and outcomes of 2023 challenge. This project is designed to challenge the computer vision community to develop robust image captioning models that advance the state-of-the-art both in terms of accuracy and fairness. Through the challenge, the image captioning models were tested using a new evaluation dataset that includes a large variety of visual concepts from many domains. There was no specific training data provided for the challenge, and therefore the challenge entries were required to adapt to new types of image descriptions that had not been seen during training. This report includes information on the newly proposed NICE dataset, evaluation methods, challenge results, and technical details of top-ranking entries. We expect that the outcomes of the challenge will contribute to the improvement of AI models on various vision-language tasks.

Dongyeun Lee, Jae Young Lee, Doyeon Kim et al.

Recent studies show strong generative performance in domain translation especially by using transfer learning techniques on the unconditional generator. However, the control between different domain features using a single model is still challenging. Existing methods often require additional models, which is computationally demanding and leads to unsatisfactory visual quality. In addition, they have restricted control steps, which prevents a smooth transition. In this paper, we propose a new approach for high-quality domain translation with better controllability. The key idea is to preserve source features within a disentangled subspace of a target feature space. This allows our method to smoothly control the degree to which it preserves source features while generating images from an entirely new domain using only a single model. Our extensive experiments show that the proposed method can produce more consistent and realistic images than previous works and maintain precise controllability over different levels of transformation. The code is available at https://github.com/LeeDongYeun/FixNoise.

Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang et al.

This survey reviews the progress of diffusion models in generating images from text, ~\textit{i.e.} text-to-image diffusion models. As a self-contained work, this survey starts with a brief introduction of how diffusion models work for image synthesis, followed by the background for text-conditioned image synthesis. Based on that, we present an organized review of pioneering methods and their improvements on text-to-image generation. We further summarize applications beyond image generation, such as text-guided generation for various modalities like videos, and text-guided image editing. Beyond the progress made so far, we discuss existing challenges and promising future directions.

Gyojin Han, Jaehyun Choi, Haeil Lee et al.

Model inversion attacks are a type of privacy attack that reconstructs private data used to train a machine learning model, solely by accessing the model. Recently, white-box model inversion attacks leveraging Generative Adversarial Networks (GANs) to distill knowledge from public datasets have been receiving great attention because of their excellent attack performance. On the other hand, current black-box model inversion attacks that utilize GANs suffer from issues such as being unable to guarantee the completion of the attack process within a predetermined number of query accesses or achieve the same level of performance as white-box attacks. To overcome these limitations, we propose a reinforcement learning-based black-box model inversion attack. We formulate the latent space search as a Markov Decision Process (MDP) problem and solve it with reinforcement learning. Our method utilizes the confidence scores of the generated images to provide rewards to an agent. Finally, the private data can be reconstructed using the latent vectors found by the agent trained in the MDP. The experiment results on various datasets and models demonstrate that our attack successfully recovers the private information of the target model by achieving state-of-the-art attack performance. We emphasize the importance of studies on privacy-preserving machine learning by proposing a more advanced black-box model inversion attack.

Gyojin Han, Jiwan Hur, Jaehyun Choi et al.

Recent developments in 3D shape representation opened new possibilities for generating detailed 3D shapes. Despite these advances, there are few studies dealing with the generation of 4D dynamic shapes that have the form of 3D objects deforming over time. To bridge this gap, we focus on generating 4D dynamic shapes with an emphasis on both generation quality and efficiency in this paper. HyperDiffusion, a previous work on 4D generation, proposed a method of directly generating the weight parameters of 4D occupancy fields but suffered from low temporal consistency and slow rendering speed due to motion representation that is not separated from the shape representation of 4D occupancy fields. Therefore, we propose a new neural deformation representation and combine it with conditional neural signed distance fields to design a 4D representation architecture in which the motion latent space is disentangled from the shape latent space. The proposed deformation representation, which works by predicting skinning weights and rigid transformations for multiple parts, also has advantages over the deformation modules of existing 4D representations in understanding the structure of shapes. In addition, we design a training process of a diffusion model that utilizes the shape and motion features that are extracted by our 4D representation as data points. The results of unconditional generation, conditional generation, and motion retargeting experiments demonstrate that our method not only shows better performance than previous works in 4D dynamic shape generation but also has various potential applications.

Chanho Lee, Seunghee Koh, Yunho Jeon et al.

The DEtection TRansformer (DETR) is a powerful end-to-end object detector, yet its one-to-one matching strategy suffers from slow convergence and low recall. A common approach to address this issue is to use one-to-many label assignment to provide more positive samples. However, existing methods that use one-to-many matching as an auxiliary objective lead to increased training costs, with their auxiliary decoders discarded during inference. To address this limitation, we propose MDS-DETR, which leverages both one-to-one and one-to-many supervision within a single decoder. Specifically, we introduce a Masked Duplicate Suppressor (MDS) that injects asymmetry into self-attention via confidence-based causal masking. MDS filters out the duplicates generated by the one-to-many supervised layer, enables explainable, duplicate-free predictions in a fully end-to-end framework. MDS-DETR outperforms existing one-to-many DETR variants such as MS-DETR, MR.DETR and Relation-DETR, without relying on any additional queries or auxiliary decoders. Under a 12-epoch training schedule on MS COCO with a ResNet-50 backbone, MDS-DETR achieves a +2.8 mAP improvement over Deformable-DETR with only a 5\% increase in training time, and outperforms the state-of-the-art MR.DETR by +0.3 mAP while being even 20\% faster in training. Our code and models are available at \href{https://github.com/dcholee/mds-detr}{https://github.com/DChoLee/MDS-DETR}.

Gyojin Han, Junmo Kim

We address text-based 3D human motion editing, where the goal is to preserve the style and structure of a source motion while applying edits described in natural language. The release of the MotionFix dataset has spurred active research into training-based diffusion models that directly generate an edited motion from a source motion and a text instruction. While previous works have focused primarily on learning when an edit should occur temporally, our goal is to create a model that understands not only this temporal aspect but also which specific joints are responsible for the change. Targeting this, we propose a novel architecture and a complementary auxiliary task to aid its training. Our architecture consists of two axis-anchored transformers, which extract distinct features along the joint and time dimensions respectively, and a cross-axis fusion block that integrates these representations. We further introduce an auxiliary task that trains the joint-anchored transformer to regress the Soft-DTW distance between source and target joint rotations. This objective teaches the module to understand which joints to modify and which to preserve. Through comprehensive experiments on the MotionFix dataset, we demonstrate that our method significantly improves semantic alignment with both the text instruction and the source motion, as well as the overall fidelity of the generated motion, achieving state-of-the-art results.

Hyungmin Kim, Sungho Suh, Daehwan Kim et al.

Recent advances in deep learning have significantly improved the performance of various computer vision applications. However, discovering novel categories in an incremental learning scenario remains a challenging problem due to the lack of prior knowledge about the number and nature of new categories. Existing methods for novel category discovery are limited by their reliance on labeled datasets and prior knowledge about the number of novel categories and the proportion of novel samples in the batch. To address the limitations and more accurately reflect real-world scenarios, in this paper, we propose a novel unsupervised class incremental learning approach for discovering novel categories on unlabeled sets without prior knowledge. The proposed method fine-tunes the feature extractor and proxy anchors on labeled sets, then splits samples into old and novel categories and clusters on the unlabeled dataset. Furthermore, the proxy anchors-based exemplar generates representative category vectors to mitigate catastrophic forgetting. Experimental results demonstrate that our proposed approach outperforms the state-of-the-art methods on fine-grained datasets under real-world scenarios.

Hanbyel Cho, Yooshin Cho, Jaesung Ahn et al.

From an image of a person, we can easily infer the natural 3D pose and shape of the person even if ambiguity exists. This is because we have a mental model that allows us to imagine a person's appearance at different viewing directions from a given image and utilize the consistency between them for inference. However, existing human mesh recovery methods only consider the direction in which the image was taken due to their structural limitations. Hence, we propose "Implicit 3D Human Mesh Recovery (ImpHMR)" that can implicitly imagine a person in 3D space at the feature-level via Neural Feature Fields. In ImpHMR, feature fields are generated by CNN-based image encoder for a given image. Then, the 2D feature map is volume-rendered from the feature field for a given viewing direction, and the pose and shape parameters are regressed from the feature. To utilize consistency with pose and shape from unseen-view, if there are 3D labels, the model predicts results including the silhouette from an arbitrary direction and makes it equal to the rotated ground-truth. In the case of only 2D labels, we perform self-supervised learning through the constraint that the pose and shape parameters inferred from different directions should be the same. Extensive evaluations show the efficacy of the proposed method.

Hyounguk Shon, Janghyeon Lee, Seung Hwan Kim et al.

Pre-trained representation is one of the key elements in the success of modern deep learning. However, existing works on continual learning methods have mostly focused on learning models incrementally from scratch. In this paper, we explore an alternative framework to incremental learning where we continually fine-tune the model from a pre-trained representation. Our method takes advantage of linearization technique of a pre-trained neural network for simple and effective continual learning. We show that this allows us to design a linear model where quadratic parameter regularization method is placed as the optimal continual learning policy, and at the same time enjoying the high performance of neural networks. We also show that the proposed algorithm enables parameter regularization methods to be applied to class-incremental problems. Additionally, we provide a theoretical reason why the existing parameter-space regularization algorithms such as EWC underperform on neural networks trained with cross-entropy loss. We show that the proposed method can prevent forgetting while achieving high continual fine-tuning performance on image classification tasks. To show that our method can be applied to general continual learning settings, we evaluate our method in data-incremental, task-incremental, and class-incremental learning problems.

Jiwan Hur, Jae Young Lee, Jaehyun Choi et al.

Light field (LF) camera captures rich information from a scene. Using the information, the LF de-occlusion (LF-DeOcc) task aims to reconstruct the occlusion-free center view image. Existing LF-DeOcc studies mainly focus on the sparsely sampled (sparse) LF images where most of the occluded regions are visible in other views due to the large disparity. In this paper, we expand LF-DeOcc in more challenging datasets, densely sampled (dense) LF images, which are taken by a micro-lens-based portable LF camera. Due to the small disparity ranges of dense LF images, most of the background regions are invisible in any view. To apply LF-DeOcc in both LF datasets, we propose a framework, ISTY, which is defined and divided into three roles: (1) extract LF features, (2) define the occlusion, and (3) inpaint occluded regions. By dividing the framework into three specialized components according to the roles, the development and analysis can be easier. Furthermore, an explainable intermediate representation, an occlusion mask, can be obtained in the proposed framework. The occlusion mask is useful for comprehensive analysis of the model and other applications by manipulating the mask. In experiments, qualitative and quantitative results show that the proposed framework outperforms state-of-the-art LF-DeOcc methods in both sparse and dense LF datasets.

Hyungmin Kim, Sungho Suh, Sunghyun Baek et al.

We present a novel adversarial penalized self-knowledge distillation method, named adversarial learning and implicit regularization for self-knowledge distillation (AI-KD), which regularizes the training procedure by adversarial learning and implicit distillations. Our model not only distills the deterministic and progressive knowledge which are from the pre-trained and previous epoch predictive probabilities but also transfers the knowledge of the deterministic predictive distributions using adversarial learning. The motivation is that the self-knowledge distillation methods regularize the predictive probabilities with soft targets, but the exact distributions may be hard to predict. Our method deploys a discriminator to distinguish the distributions between the pre-trained and student models while the student model is trained to fool the discriminator in the trained procedure. Thus, the student model not only can learn the pre-trained model's predictive probabilities but also align the distributions between the pre-trained and student models. We demonstrate the effectiveness of the proposed method with network architectures on multiple datasets and show the proposed method achieves better performance than state-of-the-art methods.

Gyojin Han, Dong-Jae Lee, Jiwan Hur et al.

Steganography is the process of embedding secret data into another message or data, in such a way that it is not easily noticeable. With the advancement of deep learning, Deep Neural Networks (DNNs) have recently been utilized in steganography. However, existing deep steganography techniques are limited in scope, as they focus on specific data types and are not effective for cross-modal steganography. Therefore, We propose a deep cross-modal steganography framework using Implicit Neural Representations (INRs) to hide secret data of various formats in cover images. The proposed framework employs INRs to represent the secret data, which can handle data of various modalities and resolutions. Experiments on various secret datasets of diverse types demonstrate that the proposed approach is expandable and capable of accommodating different modalities.

Dong-Jae Lee, Jae Young Lee, Hyounguk Shon et al.

Depth estimation is an important task in various robotics systems and applications. In mobile robotics systems, monocular depth estimation is desirable since a single RGB camera can be deployable at a low cost and compact size. Due to its significant and growing needs, many lightweight monocular depth estimation networks have been proposed for mobile robotics systems. While most lightweight monocular depth estimation methods have been developed using convolution neural networks, the Transformer has been gradually utilized in monocular depth estimation recently. However, massive parameters and large computational costs in the Transformer disturb the deployment to embedded devices. In this paper, we present a Token-Sharing Transformer (TST), an architecture using the Transformer for monocular depth estimation, optimized especially in embedded devices. The proposed TST utilizes global token sharing, which enables the model to obtain an accurate depth prediction with high throughput in embedded devices. Experimental results show that TST outperforms the existing lightweight monocular depth estimation methods. On the NYU Depth v2 dataset, TST can deliver depth maps up to 63.4 FPS in NVIDIA Jetson nano and 142.6 FPS in NVIDIA Jetson TX2, with lower errors than the existing methods. Furthermore, TST achieves real-time depth estimation of high-resolution images on Jetson TX2 with competitive results.

Gyojin Han, Jaehyun Choi, Hyeong Gwon Hong et al.

Generally, regularization-based continual learning models limit access to the previous task data to imitate the real-world constraints related to memory and privacy. However, this introduces a problem in these models by not being able to track the performance on each task. In essence, current continual learning methods are susceptible to attacks on previous tasks. We demonstrate the vulnerability of regularization-based continual learning methods by presenting a simple task-specific data poisoning attack that can be used in the learning process of a new task. Training data generated by the proposed attack causes performance degradation on a specific task targeted by the attacker. We experiment with the attack on the two representative regularization-based continual learning methods, Elastic Weight Consolidation (EWC) and Synaptic Intelligence (SI), trained with variants of MNIST dataset. The experiment results justify the vulnerability proposed in this paper and demonstrate the importance of developing continual learning models that are robust to adversarial attacks.

Jinhyung Kim, Taeoh Kim, Minho Shim et al.

Recent self-supervised video representation learning methods focus on maximizing the similarity between multiple augmented views from the same video and largely rely on the quality of generated views. However, most existing methods lack a mechanism to prevent representation learning from bias towards static information in the video. In this paper, we propose frequency augmentation (FreqAug), a spatio-temporal data augmentation method in the frequency domain for video representation learning. FreqAug stochastically removes specific frequency components from the video so that learned representation captures essential features more from the remaining information for various downstream tasks. Specifically, FreqAug pushes the model to focus more on dynamic features rather than static features in the video via dropping spatial or temporal low-frequency components. To verify the generality of the proposed method, we experiment with FreqAug on multiple self-supervised learning frameworks along with standard augmentations. Transferring the improved representation to five video action recognition and two temporal action localization downstream tasks shows consistent improvements over baselines.

Yooshin Cho, Youngsoo Kim, Hanbyel Cho et al.

Non-local (NL) block is a popular module that demonstrates the capability to model global contexts. However, NL block generally has heavy computation and memory costs, so it is impractical to apply the block to high-resolution feature maps. In this paper, to investigate the efficacy of NL block, we empirically analyze if the magnitude and direction of input feature vectors properly affect the attention between vectors. The results show the inefficacy of softmax operation which is generally used to normalize the attention map of the NL block. Attention maps normalized with softmax operation highly rely upon magnitude of key vectors, and performance is degenerated if the magnitude information is removed. By replacing softmax operation with the scaling factor, we demonstrate improved performance on CIFAR-10, CIFAR-100, and Tiny-ImageNet. In Addition, our method shows robustness to embedding channel reduction and embedding weight initialization. Notably, our method makes multi-head attention employable without additional computational cost.

Jongsuk Kim, Jiwon Shin, Junmo Kim

In recent years, advancements in representation learning and language models have propelled Automated Captioning (AC) to new heights, enabling the generation of human-level descriptions. Leveraging these advancements, we propose AVCap, an Audio-Visual Captioning framework, a simple yet powerful baseline approach applicable to audio-visual captioning. AVCap utilizes audio-visual features as text tokens, which has many advantages not only in performance but also in the extensibility and scalability of the model. AVCap is designed around three pivotal dimensions: the exploration of optimal audio-visual encoder architectures, the adaptation of pre-trained models according to the characteristics of generated text, and the investigation into the efficacy of modality fusion in captioning. Our method outperforms existing audio-visual captioning methods across all metrics and the code is available on https://github.com/JongSuk1/AVCap

Seunghee Koh, Hyounguk Shon, Janghyeon Lee et al.

Transfer learning is widely used for training deep neural networks (DNN) for building a powerful representation. Even after the pre-trained model is adapted for the target task, the representation performance of the feature extractor is retained to some extent. As the performance of the pre-trained model can be considered the private property of the owner, it is natural to seek the exclusive right of the generalized performance of the pre-trained weight. To address this issue, we suggest a new paradigm of transfer learning called disposable transfer learning (DTL), which disposes of only the source task without degrading the performance of the target task. To achieve knowledge disposal, we propose a novel loss named Gradient Collision loss (GC loss). GC loss selectively unlearns the source knowledge by leading the gradient vectors of mini-batches in different directions. Whether the model successfully unlearns the source task is measured by piggyback learning accuracy (PL accuracy). PL accuracy estimates the vulnerability of knowledge leakage by retraining the scrubbed model on a subset of source data or new downstream data. We demonstrate that GC loss is an effective approach to the DTL problem by showing that the model trained with GC loss retains the performance on the target task with a significantly reduced PL accuracy.

Dongyeun Lee, Jae Young Lee, Doyeon Kim et al.

Transfer learning of StyleGAN has recently shown great potential to solve diverse tasks, especially in domain translation. Previous methods utilized a source model by swapping or freezing weights during transfer learning, however, they have limitations on visual quality and controlling source features. In other words, they require additional models that are computationally demanding and have restricted control steps that prevent a smooth transition. In this paper, we propose a new approach to overcome these limitations. Instead of swapping or freezing, we introduce a simple feature matching loss to improve generation quality. In addition, to control the degree of source features, we train a target model with the proposed strategy, FixNoise, to preserve the source features only in a disentangled subspace of a target feature space. Owing to the disentangled feature space, our method can smoothly control the degree of the source features in a single model. Extensive experiments demonstrate that the proposed method can generate more consistent and realistic images than previous works.

Haeil Lee, Hansang Lee, Junmo Kim

Mixed Sample Data Augmentation (MSDA) techniques, such as Mixup, CutMix, and PuzzleMix, have been widely acknowledged for enhancing performance in a variety of tasks. A previous study reported the class dependency of traditional data augmentation (DA), where certain classes benefit disproportionately compared to others. This paper reveals a class dependent effect of MSDA, where some classes experience improved performance while others experience degraded performance. This research addresses the issue of class dependency in MSDA and proposes an algorithm to mitigate it. The approach involves training on a mixture of MSDA and non-MSDA data, which not only mitigates the negative impact on the affected classes, but also improves overall accuracy. Furthermore, we provide in-depth analysis and discussion of why MSDA introduced class dependencies and which classes are most likely to have them.

Eojindl Yi, Juyoung Yang, Junmo Kim

Despite its importance, unsupervised domain adaptation (UDA) on LiDAR semantic segmentation is a task that has not received much attention from the research community. Only recently, a completion-based 3D method has been proposed to tackle the problem and formally set up the adaptive scenarios. However, the proposed pipeline is complex, voxel-based and requires multi-stage inference, which inhibits it for real-time inference. We propose a range image-based, effective and efficient method for solving UDA on LiDAR segmentation. The method exploits class prototypes from the source domain to pseudo label target domain pixels, which is a research direction showing good performance in UDA for natural image semantic segmentation. Applying such approaches to LiDAR scans has not been considered because of the severe domain shift and lack of pre-trained feature extractor that is unavailable in the LiDAR segmentation setup. However, we show that proper strategies, including reconstruction-based pre-training, enhanced prototypes, and selective pseudo labeling based on distance to prototypes, is sufficient enough to enable the use of prototypical approaches. We evaluate the performance of our method on the recently proposed LiDAR segmentation UDA scenarios. Our method achieves remarkable performance among contemporary methods.

Hansang Lee, Haeil Lee, Helen Hong et al.

As the size of the dataset used in deep learning tasks increases, the noisy label problem, which is a task of making deep learning robust to the incorrectly labeled data, has become an important task. In this paper, we propose a method of learning noisy label data using the label noise selection with test-time augmentation (TTA) cross-entropy and classifier learning with the NoiseMix method. In the label noise selection, we propose TTA cross-entropy by measuring the cross-entropy to predict the test-time augmented training data. In the classifier learning, we propose the NoiseMix method based on MixUp and BalancedMix methods by mixing the samples from the noisy and the clean label data. In experiments on the ISIC-18 public skin lesion diagnosis dataset, the proposed TTA cross-entropy outperformed the conventional cross-entropy and the TTA uncertainty in detecting label noise data in the label noise selection process. Moreover, the proposed NoiseMix not only outperformed the state-of-the-art methods in the classification performance but also showed the most robustness to the label noise in the classifier learning.

Hansang Lee, Haeil Lee, Helen Hong et al.

Uncertainty estimation of trained deep learning networks is valuable for optimizing learning efficiency and evaluating the reliability of network predictions. In this paper, we propose a method for estimating uncertainty in deep learning image classification using test-time mixup augmentation (TTMA). To improve the ability to distinguish correct and incorrect predictions in existing aleatoric uncertainty, we introduce TTMA data uncertainty (TTMA-DU) by applying mixup augmentation to test data and measuring the entropy of the predicted label histogram. In addition to TTMA-DU, we propose TTMA class-specific uncertainty (TTMA-CSU), which captures aleatoric uncertainty specific to individual classes and provides insight into class confusion and class similarity within the trained network. We validate our proposed methods on the ISIC-18 skin lesion diagnosis dataset and the CIFAR-100 real-world image classification dataset. Our experiments show that (1) TTMA-DU more effectively differentiates correct and incorrect predictions compared to existing uncertainty measures due to mixup perturbation, and (2) TTMA-CSU provides information on class confusion and class similarity for both datasets.

Hanbyel Cho, Yekang Lee, Jaemyung Yu et al.

When a high-resolution (HR) image is degraded into a low-resolution (LR) image, the image loses some of the existing information. Consequently, multiple HR images can correspond to the LR image. Most of the existing methods do not consider the uncertainty caused by the stochastic attribute, which can only be probabilistically inferred. Therefore, the predicted HR images are often blurry because the network tries to reflect all possibilities in a single output image. To overcome this limitation, this paper proposes a novel face super-resolution (SR) scheme to take into the uncertainty by stochastic modeling. Specifically, the information in LR images is separately encoded into deterministic and stochastic attributes. Furthermore, an Input Conditional Attribute Predictor is proposed and separately trained to predict the partially alive stochastic attributes from only the LR images. Extensive evaluation shows that the proposed method successfully reduces the uncertainty in the learning process and outperforms the existing state-of-the-art approaches.

Youngwoo Shin, Jiwan Hur, Junmo Kim

Visual autoregressive (VAR) models generate images through next-scale prediction, naturally achieving coarse-to-fine, fast, high-fidelity synthesis mirroring human perception. In practice, this hierarchy can drift at inference time, as limited capacity and accumulated error cause the model to deviate from its coarse-to-fine nature. We revisit this limitation from an information-theoretic perspective and deduce that ensuring each scale contributes high-frequency content not explained by earlier scales mitigates the train-inference discrepancy. With this insight, we propose Scaled Spatial Guidance (SSG), training-free, inference-time guidance that steers generation toward the intended hierarchy while maintaining global coherence. SSG emphasizes target high-frequency signals, defined as the semantic residual, isolated from a coarser prior. To obtain this prior, we leverage a principled frequency-domain procedure, Discrete Spatial Enhancement (DSE), which is devised to sharpen and better isolate the semantic residual through frequency-aware construction. SSG applies broadly across VAR models leveraging discrete visual tokens, regardless of tokenization design or conditioning modality. Experiments demonstrate SSG yields consistent gains in fidelity and diversity while preserving low latency, revealing untapped efficiency in coarse-to-fine image generation. Code is available at https://github.com/Youngwoo-git/SSG.

Haeil Lee, Hansang Lee, Seoyeon Gye et al.

Generative diffusion models have emerged as a powerful tool for high-quality image synthesis, yet their iterative nature demands significant computational resources. This paper proposes an efficient time step sampling method based on an image spectral analysis of the diffusion process, aimed at optimizing the denoising process. Instead of the traditional uniform distribution-based time step sampling, we introduce a Beta distribution-like sampling technique that prioritizes critical steps in the early and late stages of the process. Our hypothesis is that certain steps exhibit significant changes in image content, while others contribute minimally. We validated our approach using Fourier transforms to measure frequency response changes at each step, revealing substantial low-frequency changes early on and high-frequency adjustments later. Experiments with ADM and Stable Diffusion demonstrated that our Beta Sampling method consistently outperforms uniform sampling, achieving better FID and IS scores, and offers competitive efficiency relative to state-of-the-art methods like AutoDiffusion. This work provides a practical framework for enhancing diffusion model efficiency by focusing computational resources on the most impactful steps, with potential for further optimization and broader application.

Chenshuang Zhang, Fei Pan, Junmo Kim et al.

We establish rigorous benchmarks for visual perception robustness. Synthetic images such as ImageNet-C, ImageNet-9, and Stylized ImageNet provide specific type of evaluation over synthetic corruptions, backgrounds, and textures, yet those robustness benchmarks are restricted in specified variations and have low synthetic quality. In this work, we introduce generative model as a data source for synthesizing hard images that benchmark deep models' robustness. Leveraging diffusion models, we are able to generate images with more diversified backgrounds, textures, and materials than any prior work, where we term this benchmark as ImageNet-D. Experimental results show that ImageNet-D results in a significant accuracy drop to a range of vision models, from the standard ResNet visual classifier to the latest foundation models like CLIP and MiniGPT-4, significantly reducing their accuracy by up to 60\%. Our work suggests that diffusion models can be an effective source to test vision models. The code and dataset are available at https://github.com/chenshuang-zhang/imagenet_d.

Jiwan Hur, Jaehyun Choi, Gyojin Han et al.

Training diffusion models on limited datasets poses challenges in terms of limited generation capacity and expressiveness, leading to unsatisfactory results in various downstream tasks utilizing pretrained diffusion models, such as domain translation and text-guided image manipulation. In this paper, we propose Self-Distillation for Fine-Tuning diffusion models (SDFT), a methodology to address these challenges by leveraging diverse features from diffusion models pretrained on large source datasets. SDFT distills more general features (shape, colors, etc.) and less domain-specific features (texture, fine details, etc) from the source model, allowing successful knowledge transfer without disturbing the training process on target datasets. The proposed method is not constrained by the specific architecture of the model and thus can be generally adopted to existing frameworks. Experimental results demonstrate that SDFT enhances the expressiveness of the diffusion model with limited datasets, resulting in improved generation capabilities across various downstream tasks.

Myungchul Kim, Kwanyong Park, Junmo Kim et al.

We introduce ARGOS, the first benchmark and framework that reformulates multi-camera person search as an interactive reasoning problem requiring an agent to plan, question, and eliminate candidates under information asymmetry. An ARGOS agent receives a vague witness statement and must decide what to ask, when to invoke spatial or temporal tools, and how to interpret ambiguous responses, all within a limited turn budget. Reasoning is grounded in a Spatio-Temporal Topology Graph (STTG) encoding camera connectivity and empirically validated transition times. The benchmark comprises 2,691 tasks across 14 real-world scenarios in three progressive tracks: semantic perception (Who), spatial reasoning (Where), and temporal reasoning (When). Experiments with four LLM backbones show the benchmark is far from solved (best TWS: 0.383 on Track 2, 0.590 on Track 3), and ablations confirm that removing domain-specific tools drops accuracy by up to 49.6 percentage points.

Hoang C. Nguyen, Haeil Lee, Junmo Kim

Transformer becomes more popular in the vision domain in recent years so there is a need for finding an effective way to interpret the Transformer model by visualizing it. In recent work, Chefer et al. can visualize the Transformer on vision and multi-modal tasks effectively by combining attention layers to show the importance of each image patch. However, when applying to other variants of Transformer such as the Swin Transformer, this method can not focus on the predicted object. Our method, by considering the statistics of tokens in layer normalization layers, shows a great ability to interpret the explainability of Swin Transformer and ViT.

Chenshuang Zhang, Kang Zhang, Joon Son Chung et al.

Distinguishing visually similar objects by their motion remains a critical challenge in computer vision. Although supervised trackers show promise, contemporary self-supervised trackers struggle when visual cues become ambiguous, limiting their scalability and generalization without extensive labeled data. We find that pre-trained video diffusion models inherently learn motion representations suitable for tracking without task-specific training. This ability arises because their denoising process isolates motion in early, high-noise stages, distinct from later appearance refinement. Capitalizing on this discovery, our self-supervised tracker significantly improves performance in distinguishing visually similar objects, an underexplored failure point for existing methods. Our method achieves up to a 6-point improvement over recent self-supervised approaches on established benchmarks and our newly introduced tests focused on tracking visually similar items. Visualizations confirm that these diffusion-derived motion representations enable robust tracking of even identical objects across challenging viewpoint changes and deformations.

Seunghee Koh, Sunghyun Baek, Youngdong Kim et al.

Unlearning in large language models (LLMs) has emerged as a promising safeguard against adversarial behaviors. When the forgetting loss is applied uniformly without considering token-level semantic importance, model utility can be unnecessarily degraded. Recent studies have explored token-wise loss regularizers that prioritize informative tokens, but largely rely on ground-truth confidence or external linguistic parsers, which limits their ability to capture contextual information or the model's overall predictive state. Intuitively, function words like "the" primarily serve syntactic roles and are highly predictable with little ambiguity, but informative words admit multiple plausible alternatives with greater uncertainty. Based on this intuition, we propose Entropy-guided Token Weighting (ETW), a token-level unlearning regularizer that uses entropy of the predictive distribution as a proxy for token informativeness. We demonstrate that informative tokens tend to have higher entropy, whereas structural tokens tend to have lower entropy. This behavior enables ETW to achieve more effective unlearning while better preserving model utility than existing token-level approaches.

Minchan Kwon, Hyounguk Shon, Junmo Kim

Large multimodal models (LMMs) have recently demonstrated remarkable performance in video question answering (VideoQA), yet reasoning over video remains challenging due to high inference cost and diluted information. Keyframe selection offers efficiency and sharper reasoning but suffers from sparse supervision and redundant frame choices when relying only on image-text similarity. We present a question-aware keyframe selection framework with two components: pseudo keyframe labels derived from LMMs that provide informative supervision and a coverage regularization that promotes diverse, complementary evidence across time. Experiments on NExT-QA show that our method significantly improves accuracy, especially for temporal and causal question types, establishing keyframe selection as an effective and learnable module for VideoQA.

Sunghyun Baek, Jaemyung Yu, Seunghee Koh et al.

Test-time adaptation (TTA) has been widely explored to prevent performance degradation when test data differ from the training distribution. However, fully leveraging the rich representations of large pretrained models with minimal parameter updates remains underexplored. In this paper, we propose Intrinsic Mixture of Spectral Experts (IMSE) that leverages the spectral experts inherently embedded in Vision Transformers. We decompose each linear layer via singular value decomposition (SVD) and adapt only the singular values, while keeping the singular vectors fixed. We further identify a key limitation of entropy minimization in TTA: it often induces feature collapse, causing the model to rely on domain-specific features rather than class-discriminative features. To address this, we propose a diversity maximization loss based on expert-input alignment, which encourages diverse utilization of spectral experts during adaptation. In the continual test-time adaptation (CTTA) scenario, beyond preserving pretrained knowledge, it is crucial to retain and reuse knowledge from previously observed domains. We introduce Domain-Aware Spectral Code Retrieval, which estimates input distributions to detect domain shifts, and retrieves adapted singular values for rapid adaptation. Consequently, our method achieves state-of-the-art performance on various distribution-shift benchmarks under the TTA setting. In CTTA and Gradual CTTA, it further improves accuracy by 3.4 percentage points (pp) and 2.4 pp, respectively, while requiring 385 times fewer trainable parameters. Our code is available at https://github.com/baek85/IMSE.

Dong-Jae Lee, Jiwan Hur, Jaehyun Choi et al.

Vision Transformers have demonstrated exceptional performance across various computer vision tasks, yet their quadratic computational complexity concerning token length remains a significant challenge. To address this, token reduction methods have been widely explored. However, existing approaches often overlook the frequency characteristics of self-attention, such as rank collapsing and over-smoothing phenomenon. In this paper, we propose a frequency-aware token reduction strategy that improves computational efficiency while preserving performance by mitigating rank collapsing. Our method partitions tokens into high-frequency tokens and low-frequency tokens. high-frequency tokens are selectively preserved, while low-frequency tokens are aggregated into a compact direct current token to retain essential low-frequency components. Through extensive experiments and analysis, we demonstrate that our approach significantly improves accuracy while reducing computational overhead and mitigating rank collapsing and over smoothing. Furthermore, we analyze the previous methods, shedding light on their implicit frequency characteristics and limitations.

Dongyeun Lee, Jiwan Hur, Hyounguk Shon et al.

Diffusion models have achieved remarkable success in image generation but come with significant computational costs, posing challenges for deployment in resource-constrained environments. Recent post-training quantization (PTQ) methods have attempted to mitigate this issue by focusing on the iterative nature of diffusion models. However, these approaches often overlook outliers, leading to degraded performance at low bit-widths. In this paper, we propose a DMQ which combines Learned Equivalent Scaling (LES) and channel-wise Power-of-Two Scaling (PTS) to effectively address these challenges. Learned Equivalent Scaling optimizes channel-wise scaling factors to redistribute quantization difficulty between weights and activations, reducing overall quantization error. Recognizing that early denoising steps, despite having small quantization errors, crucially impact the final output due to error accumulation, we incorporate an adaptive timestep weighting scheme to prioritize these critical steps during learning. Furthermore, identifying that layers such as skip connections exhibit high inter-channel variance, we introduce channel-wise Power-of-Two Scaling for activations. To ensure robust selection of PTS factors even with small calibration set, we introduce a voting algorithm that enhances reliability. Extensive experiments demonstrate that our method significantly outperforms existing works, especially at low bit-widths such as W4A6 (4-bit weight, 6-bit activation) and W4A8, maintaining high image generation quality and model stability. The code is available at https://github.com/LeeDongYeun/dmq.

Jaemyung Yu, Jaehyun Choi, Dong-Jae Lee et al.

Unsupervised representation learning has significantly advanced various machine learning tasks. In the computer vision domain, state-of-the-art approaches utilize transformations like random crop and color jitter to achieve invariant representations, embedding semantically the same inputs despite transformations. However, this can degrade performance in tasks requiring precise features, such as localization or flower classification. To address this, recent research incorporates equivariant representation learning, which captures transformation-sensitive information. However, current methods depend on transformation labels and thus struggle with interdependency and complex transformations. We propose Self-supervised Transformation Learning (STL), replacing transformation labels with transformation representations derived from image pairs. The proposed method ensures transformation representation is image-invariant and learns corresponding equivariant transformations, enhancing performance without increased batch complexity. We demonstrate the approach's effectiveness across diverse classification and detection tasks, outperforming existing methods in 7 out of 11 benchmarks and excelling in detection. By integrating complex transformations like AugMix, unusable by prior equivariant methods, this approach enhances performance across tasks, underscoring its adaptability and resilience. Additionally, its compatibility with various base models highlights its flexibility and broad applicability. The code is available at https://github.com/jaemyung-u/stl.

Youngtaek Oh, Pyunghwan Ahn, Jinhyung Kim et al.

Vision and language models (VLMs) such as CLIP have showcased remarkable zero-shot recognition abilities yet face challenges in visio-linguistic compositionality, particularly in linguistic comprehension and fine-grained image-text alignment. This paper explores the intricate relationship between compositionality and recognition -- two pivotal aspects of VLM capability. We conduct a comprehensive evaluation of existing VLMs, covering both pre-training approaches aimed at recognition and the fine-tuning methods designed to improve compositionality. Our evaluation employs 12 benchmarks for compositionality, along with 21 zero-shot classification and two retrieval benchmarks for recognition. In our analysis from 274 CLIP model checkpoints, we reveal patterns and trade-offs that emerge between compositional understanding and recognition accuracy. Ultimately, this necessitates strategic efforts towards developing models that improve both capabilities, as well as the meticulous formulation of benchmarks for compositionality. We open our evaluation framework at https://github.com/ytaek-oh/vl_compo.

Jongsuk Kim, Hyeongkeun Lee, Kyeongha Rho et al.

Recent advancements in self-supervised audio-visual representation learning have demonstrated its potential to capture rich and comprehensive representations. However, despite the advantages of data augmentation verified in many learning methods, audio-visual learning has struggled to fully harness these benefits, as augmentations can easily disrupt the correspondence between input pairs. To address this limitation, we introduce EquiAV, a novel framework that leverages equivariance for audio-visual contrastive learning. Our approach begins with extending equivariance to audio-visual learning, facilitated by a shared attention-based transformation predictor. It enables the aggregation of features from diverse augmentations into a representative embedding, providing robust supervision. Notably, this is achieved with minimal computational overhead. Extensive ablation studies and qualitative results verify the effectiveness of our method. EquiAV outperforms previous works across various audio-visual benchmarks. The code is available on https://github.com/JongSuk1/EquiAV.

Pyunghwan Ahn, Juyoung Yang, Eojindl Yi et al.

Understanding point cloud has recently gained huge interests following the development of 3D scanning devices and the accumulation of large-scale 3D data. Most point cloud processing algorithms can be classified as either point-based or voxel-based methods, both of which have severe limitations in processing time or memory, or both. To overcome these limitations, we propose Projection-based Point Convolution (PPConv), a point convolutional module that uses 2D convolutions and multi-layer perceptrons (MLPs) as its components. In PPConv, point features are processed through two branches: point branch and projection branch. Point branch consists of MLPs, while projection branch transforms point features into a 2D feature map and then apply 2D convolutions. As PPConv does not use point-based or voxel-based convolutions, it has advantages in fast point cloud processing. When combined with a learnable projection and effective feature fusion strategy, PPConv achieves superior efficiency compared to state-of-the-art methods, even with a simple architecture based on PointNet++. We demonstrate the efficiency of PPConv in terms of the trade-off between inference time and segmentation performance. The experimental results on S3DIS and ShapeNetPart show that PPConv is the most efficient method among the compared ones. The code is available at github.com/pahn04/PPConv.

Beomyoung Kim, Youngjoon Yoo, Chaeeun Rhee et al.

Weakly-supervised instance segmentation (WSIS) has been considered as a more challenging task than weakly-supervised semantic segmentation (WSSS). Compared to WSSS, WSIS requires instance-wise localization, which is difficult to extract from image-level labels. To tackle the problem, most WSIS approaches use off-the-shelf proposal techniques that require pre-training with instance or object level labels, deviating the fundamental definition of the fully-image-level supervised setting. In this paper, we propose a novel approach including two innovative components. First, we propose a semantic knowledge transfer to obtain pseudo instance labels by transferring the knowledge of WSSS to WSIS while eliminating the need for the off-the-shelf proposals. Second, we propose a self-refinement method to refine the pseudo instance labels in a self-supervised scheme and to use the refined labels for training in an online manner. Here, we discover an erroneous phenomenon, semantic drift, that occurred by the missing instances in pseudo instance labels categorized as background class. This semantic drift occurs confusion between background and instance in training and consequently degrades the segmentation performance. We term this problem as semantic drift problem and show that our proposed self-refinement method eliminates the semantic drift problem. The extensive experiments on PASCAL VOC 2012 and MS COCO demonstrate the effectiveness of our approach, and we achieve a considerable performance without off-the-shelf proposal techniques. The code is available at https://github.com/clovaai/BESTIE.

Beomyoung Kim, Janghyeon Lee, Sihaeng Lee et al.

We present a novel approach for oriented object detection, named TricubeNet, which localizes oriented objects using visual cues ($i.e.,$ heatmap) instead of oriented box offsets regression. We represent each object as a 2D Tricube kernel and extract bounding boxes using simple image-processing algorithms. Our approach is able to (1) obtain well-arranged boxes from visual cues, (2) solve the angle discontinuity problem, and (3) can save computational complexity due to our anchor-free modeling. To further boost the performance, we propose some effective techniques for size-invariant loss, reducing false detections, extracting rotation-invariant features, and heatmap refinement. To demonstrate the effectiveness of our TricubeNet, we experiment on various tasks for weakly-occluded oriented object detection: detection in an aerial image, densely packed object image, and text image. The extensive experimental results show that our TricubeNet is quite effective for oriented object detection. Code is available at https://github.com/qjadud1994/TricubeNet.

Beomyoung Kim, Sangeun Han, Junmo Kim

Weakly-supervised semantic segmentation (WSSS) using image-level labels has recently attracted much attention for reducing annotation costs. Existing WSSS methods utilize localization maps from the classification network to generate pseudo segmentation labels. However, since localization maps obtained from the classifier focus only on sparse discriminative object regions, it is difficult to generate high-quality segmentation labels. To address this issue, we introduce discriminative region suppression (DRS) module that is a simple yet effective method to expand object activation regions. DRS suppresses the attention on discriminative regions and spreads it to adjacent non-discriminative regions, generating dense localization maps. DRS requires few or no additional parameters and can be plugged into any network. Furthermore, we introduce an additional learning strategy to give a self-enhancement of localization maps, named localization map refinement learning. Benefiting from this refinement learning, localization maps are refined and enhanced by recovering some missing parts or removing noise itself. Due to its simplicity and effectiveness, our approach achieves mIoU 71.4% on the PASCAL VOC 2012 segmentation benchmark using only image-level labels. Extensive experiments demonstrate the effectiveness of our approach. The code is available at https://github.com/qjadud1994/DRS.

Dongyoon Han, Jiwhan Kim, Junmo Kim

Deep convolutional neural networks (DCNNs) have shown remarkable performance in image classification tasks in recent years. Generally, deep neural network architectures are stacks consisting of a large number of convolutional layers, and they perform downsampling along the spatial dimension via pooling to reduce memory usage. Concurrently, the feature map dimension (i.e., the number of channels) is sharply increased at downsampling locations, which is essential to ensure effective performance because it increases the diversity of high-level attributes. This also applies to residual networks and is very closely related to their performance. In this research, instead of sharply increasing the feature map dimension at units that perform downsampling, we gradually increase the feature map dimension at all units to involve as many locations as possible. This design, which is discussed in depth together with our new insights, has proven to be an effective means of improving generalization ability. Furthermore, we propose a novel residual unit capable of further improving the classification accuracy with our new network architecture. Experiments on benchmark CIFAR-10, CIFAR-100, and ImageNet datasets have shown that our network architecture has superior generalization ability compared to the original residual networks. Code is available at https://github.com/jhkim89/PyramidNet}

Gayoung Lee, Junho Kim, Jin-Hwa Kim et al.

Understanding reflection remains a long-standing challenge in 3D reconstruction due to the entanglement of appearance and geometry under view-dependent reflections. In this work, we present the Pygmalion Effect in Vision, a novel framework that metaphorically "sculpts" reflective objects into clay-like forms through image-to-clay translation. Inspired by the myth of Pygmalion, our method learns to suppress specular cues while preserving intrinsic geometric consistency, enabling robust reconstruction from multi-view images containing complex reflections. Specifically, we introduce a dual-branch network in which a BRDF-based reflective branch is complemented by a clay-guided branch that stabilizes geometry and refines surface normals. The two branches are trained jointly using the synthesized clay-like images, which provide a neutral, reflection-free supervision signal that complements the reflective views. Experiments on both synthetic and real datasets demonstrate substantial improvement in normal accuracy and mesh completeness over existing reflection-handling methods. Beyond technical gains, our framework reveals that seeing by unshining, translating radiance into neutrality, can serve as a powerful inductive bias for reflective object geometry learning.

Minsu Kim, Dongyeun Lee, Jaemyung Yu et al.

Object detection is pivotal in computer vision, yet its immense computational demands make deployment slow and power-hungry, motivating quantization. However, task-irrelevant morphologies such as background clutter and sensor noise induce redundant activations (or anomalies). These anomalies expand activation ranges and skew activation distributions toward task-irrelevant responses, complicating bit allocation and weakening the preservation of informative features. Without a clear criterion to distinguish anomalies, suppressing them can inadvertently discard useful information. To address this, we present InlierQ, an inlier-centric post-training quantization approach that separates anomalies from informative inliers. InlierQ computes gradient-aware volume saliency scores, classifies each volume as an inlier or anomaly, and fits a posterior distribution over these scores using the Expectation-Maximization (EM) algorithm. This design suppresses anomalies while preserving informative features. InlierQ is label-free, drop-in, and requires only 64 calibration samples. Experiments on the COCO and nuScenes benchmarks show consistent reductions in quantization error for camera-based (2D and 3D) and LiDAR-based (3D) object detection.

Minseo Kim, Minchan Kwon, Dongyeun Lee et al.

Personalized text-to-image generation suffers from concept entanglement, where irrelevant residual information from reference images is captured, leading to a trade-off between concept fidelity and text alignment. Recent disentanglement approaches attempt to solve this utilizing manual guidance, such as linguistic cues or segmentation masks, which limits their applicability and fails to fully articulate the target concept. In this paper, we propose ConceptPrism, a novel framework that automatically disentangles the shared visual concept from image-specific residuals by comparing images within a set. Our method jointly optimizes a target token and image-wise residual tokens using two complementary objectives: a reconstruction loss to ensure fidelity, and a novel exclusion loss that compels residual tokens to discard the shared concept. This process allows the target token to capture the pure concept without direct supervision. Extensive experiments demonstrate that ConceptPrism effectively resolves concept entanglement, achieving a significantly improved trade-off between fidelity and alignment.