Taeoh Kim

Taeoh Kim, Jinhyung Kim, Minho Shim et al.

Data augmentation has recently emerged as an essential component of modern training recipes for visual recognition tasks. However, data augmentation for video recognition has been rarely explored despite its effectiveness. Few existing augmentation recipes for video recognition naively extend the image augmentation methods by applying the same operations to the whole video frames. Our main idea is that the magnitude of augmentation operations for each frame needs to be changed over time to capture the real-world video's temporal variations. These variations should be generated as diverse as possible using fewer additional hyper-parameters during training. Through this motivation, we propose a simple yet effective video data augmentation framework, DynaAugment. The magnitude of augmentation operations on each frame is changed by an effective mechanism, Fourier Sampling that parameterizes diverse, smooth, and realistic temporal variations. DynaAugment also includes an extended search space suitable for video for automatic data augmentation methods. DynaAugment experimentally demonstrates that there are additional performance rooms to be improved from static augmentations on diverse video models. Specifically, we show the effectiveness of DynaAugment on various video datasets and tasks: large-scale video recognition (Kinetics-400 and Something-Something-v2), small-scale video recognition (UCF- 101 and HMDB-51), fine-grained video recognition (Diving-48 and FineGym), video action segmentation on Breakfast, video action localization on THUMOS'14, and video object detection on MOT17Det. DynaAugment also enables video models to learn more generalized representation to improve the model robustness on the corrupted videos.

Pilhyeon Lee, Taeoh Kim, Minho Shim et al.

Temporal action detection aims to predict the time intervals and the classes of action instances in the video. Despite the promising performance, existing two-stream models exhibit slow inference speed due to their reliance on computationally expensive optical flow. In this paper, we introduce a decomposed cross-modal distillation framework to build a strong RGB-based detector by transferring knowledge of the motion modality. Specifically, instead of direct distillation, we propose to separately learn RGB and motion representations, which are in turn combined to perform action localization. The dual-branch design and the asymmetric training objectives enable effective motion knowledge transfer while preserving RGB information intact. In addition, we introduce a local attentive fusion to better exploit the multimodal complementarity. It is designed to preserve the local discriminability of the features that is important for action localization. Extensive experiments on the benchmarks verify the effectiveness of the proposed method in enhancing RGB-based action detectors. Notably, our framework is agnostic to backbones and detection heads, bringing consistent gains across different model combinations.

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.

Jinsung Lee, Taeoh Kim, Inwoong Lee et al.

Video action detection (VAD) aims to detect actors and classify their actions in a video. We figure that VAD suffers more from classification rather than localization of actors. Hence, we analyze how prevailing methods form features for classification and find that they prioritize actor regions, yet often overlooking the essential contextual information necessary for accurate classification. Accordingly, we propose to reduce the bias toward actor and encourage paying attention to the context that is relevant to each action class. By assigning a class-dedicated query to each action class, our model can dynamically determine where to focus for effective classification. The proposed model demonstrates superior performance on three challenging benchmarks with significantly fewer parameters and less computation.

Minho Shim, Taeoh Kim, Jinhyung Kim et al.

Summarizing a video requires a diverse understanding of the video, ranging from recognizing scenes to evaluating how much each frame is essential enough to be selected as a summary. Self-supervised learning (SSL) is acknowledged for its robustness and flexibility to multiple downstream tasks, but the video SSL has not shown its value for dense understanding tasks like video summarization. We claim an unsupervised autoencoder with sufficient self-supervised learning does not need any extra downstream architecture design or fine-tuning weights to be utilized as a video summarization model. The proposed method to evaluate the importance score of each frame takes advantage of the reconstruction score of the autoencoder's decoder. We evaluate the method in major unsupervised video summarization benchmarks to show its effectiveness under various experimental settings.

Seonghoon Yu, Ilchae Jung, Byeongju Han et al.

Referring image segmentation (RIS) requires dense vision-language interactions between visual pixels and textual words to segment objects based on a given description. However, commonly adapted dual-encoders in RIS, e.g., Swin transformer and BERT (uni-modal encoders) or CLIP (a multi-modal dual-encoder), lack dense multi-modal interactions during pre-training, leading to a gap with a pixel-level RIS task. To bridge this gap, existing RIS methods often rely on multi-modal fusion modules that interact two encoders, but this approach leads to high computational costs. In this paper, we present a novel RIS method with a single-encoder, i.e., BEiT-3, maximizing the potential of shared self-attention across all framework components. This enables seamless interactions of two modalities from input to final prediction, producing granularly aligned multi-modal features. Furthermore, we propose lightweight yet effective decoder modules, a Shared FPN and a Shared Mask Decoder, which contribute to the high efficiency of our model. Our simple baseline with a single encoder achieves outstanding performances on the RIS benchmark datasets while maintaining computational efficiency, compared to the most recent SoTA methods based on dual-encoders.

Geo Ahn, Inwoong Lee, Taeoh Kim et al.

We study Compositional Video Understanding (CVU), where models must recognize verbs and objects and compose them to generalize to unseen combinations. We find that existing Zero-Shot Compositional Action Recognition (ZS-CAR) models fail primarily due to an overlooked failure mode: object-driven verb shortcuts. Through systematic analysis, we show that this behavior arises from two intertwined factors: severe sparsity and skewness of compositional supervision, and the asymmetric learning difficulty between verbs and objects. As training progresses, the existing ZS-CAR model increasingly ignores visual evidence and overfits to co-occurrence statistics. Consequently, the existing model does not gain the benefit of compositional recognition in unseen verb-object compositions. To address this, we propose RCORE, a simple and effective framework that enforces temporally grounded verb learning. RCORE introduces (i) a composition-aware augmentation that diversifies verb-object combinations without corrupting motion cues, and (ii) a temporal order regularization loss that penalizes shortcut behaviors by explicitly modeling temporal structure. Across two benchmarks, Sth-com and our newly constructed EK100-com, RCORE significantly improves unseen composition accuracy, reduces reliance on co-occurrence bias, and achieves consistently positive compositional gaps. Our findings reveal object-driven shortcuts as a critical limiting factor in ZS-CAR and demonstrate that addressing them is essential for robust compositional video understanding.

Geuntaek Lim, Minho Shim, Sungjune Park et al.

The inherent complexity of video understanding makes it difficult to attribute whether performance gains stem from visual perception, linguistic reasoning, or knowledge priors. While many benchmarks have emerged to assess high-level reasoning, the essential criteria that constitute video understanding remain largely overlooked. Instead of introducing yet another benchmark, we take a step back to re-examine the current landscape of video understanding. In this work, we provide Video-Oasis, a sustainable diagnostic suite designed to systematically evaluate existing evaluations and distill spatio-temporal challenges for video understanding. Our analysis reveals two critical findings: (1) 54% of existing benchmark samples are solvable without visual input or temporal context, and (2) on the remaining samples, state-of-the-art models exhibit performance barely exceeding random guessing. To bridge this gap, we investigate which algorithmic design choices contribute to robust video understanding, providing practical guidelines for future research. We hope our work serves as a standard guideline for benchmark construction and the rigorous evaluation of architecture development. Code is available at https://github.com/sejong-rcv/Video-Oasis.

Taeoh Kim, Hyeongmin Lee, MyeongAh Cho et al.

Deep-Learning-based video recognition has shown promising improvements along with the development of large-scale datasets and spatiotemporal network architectures. In image recognition, learning spatially invariant features is a key factor in improving recognition performance and robustness. Data augmentation based on visual inductive priors, such as cropping, flipping, rotating, or photometric jittering, is a representative approach to achieve these features. Recent state-of-the-art recognition solutions have relied on modern data augmentation strategies that exploit a mixture of augmentation operations. In this study, we extend these strategies to the temporal dimension for videos to learn temporally invariant or temporally localizable features to cover temporal perturbations or complex actions in videos. Based on our novel temporal data augmentation algorithms, video recognition performances are improved using only a limited amount of training data compared to the spatial-only data augmentation algorithms, including the 1st Visual Inductive Priors (VIPriors) for data-efficient action recognition challenge. Furthermore, learned features are temporally localizable that cannot be achieved using spatial augmentation algorithms. Our source code is available at https://github.com/taeoh-kim/temporal_data_augmentation.

Jeongseok Hyun, Sukjun Hwang, Su Ho Han et al.

Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2$\times$ speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3$\times$ speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.

WonJun Moon, Cheol-Ho Cho, Woojin Jun et al.

In a retrieval system, simultaneously achieving search accuracy and efficiency is inherently challenging. This challenge is particularly pronounced in partially relevant video retrieval (PRVR), where incorporating more diverse context representations at varying temporal scales for each video enhances accuracy but increases computational and memory costs. To address this dichotomy, we propose a prototypical PRVR framework that encodes diverse contexts within a video into a fixed number of prototypes. We then introduce several strategies to enhance text association and video understanding within the prototypes, along with an orthogonal objective to ensure that the prototypes capture a diverse range of content. To keep the prototypes searchable via text queries while accurately encoding video contexts, we implement cross- and uni-modal reconstruction tasks. The cross-modal reconstruction task aligns the prototypes with textual features within a shared space, while the uni-modal reconstruction task preserves all video contexts during encoding. Additionally, we employ a video mixing technique to provide weak guidance to further align prototypes and associated textual representations. Extensive evaluations on TVR, ActivityNet-Captions, and QVHighlights validate the effectiveness of our approach without sacrificing efficiency.

Jungho Lee, Suhwan Cho, Taeoh Kim et al.

3D Gaussian Splatting (3DGS) has attracted significant attention for its high-quality novel view rendering, inspiring research to address real-world challenges. While conventional methods depend on sharp images for accurate scene reconstruction, real-world scenarios are often affected by defocus blur due to finite depth of field, making it essential to account for realistic 3D scene representation. In this study, we propose CoCoGaussian, a Circle of Confusion-aware Gaussian Splatting that enables precise 3D scene representation using only defocused images. CoCoGaussian addresses the challenge of defocus blur by modeling the Circle of Confusion (CoC) through a physically grounded approach based on the principles of photographic defocus. Exploiting 3D Gaussians, we compute the CoC diameter from depth and learnable aperture information, generating multiple Gaussians to precisely capture the CoC shape. Furthermore, we introduce a learnable scaling factor to enhance robustness and provide more flexibility in handling unreliable depth in scenes with reflective or refractive surfaces. Experiments on both synthetic and real-world datasets demonstrate that CoCoGaussian achieves state-of-the-art performance across multiple benchmarks.

Jungho Lee, Donghyeong Kim, Dogyoon Lee et al.

3D Gaussian Splatting (3DGS) has gained significant attention due to its high-quality novel view rendering, motivating research to address real-world challenges. A critical issue is the camera motion blur caused by movement during exposure, which hinders accurate 3D scene reconstruction. In this study, we propose CoMoGaussian, a Continuous Motion-Aware Gaussian Splatting that reconstructs precise 3D scenes from motion-blurred images while maintaining real-time rendering speed. Considering the complex motion patterns inherent in real-world camera movements, we predict continuous camera trajectories using neural ordinary differential equations (ODEs). To ensure accurate modeling, we employ rigid body transformations, preserving the shape and size of the object but rely on the discrete integration of sampled frames. To better approximate the continuous nature of motion blur, we introduce a continuous motion refinement (CMR) transformation that refines rigid transformations by incorporating additional learnable parameters. By revisiting fundamental camera theory and leveraging advanced neural ODE techniques, we achieve precise modeling of continuous camera trajectories, leading to improved reconstruction accuracy. Extensive experiments demonstrate state-of-the-art performance both quantitatively and qualitatively on benchmark datasets, which include a wide range of motion blur scenarios, from moderate to extreme blur.

Chajin Shin, Taeoh Kim, Sangjin Lee et al.

Deep learning-based image inpainting algorithms have shown great performance via powerful learned prior from the numerous external natural images. However, they show unpleasant results on the test image whose distribution is far from the that of training images because their models are biased toward the training images. In this paper, we propose a simple image inpainting algorithm with test-time adaptation named AdaFill. Given a single out-of-distributed test image, our goal is to complete hole region more naturally than the pre-trained inpainting models. To achieve this goal, we treat remained valid regions of the test image as another training cues because natural images have strong internal similarities. From this test-time adaptation, our network can exploit externally learned image priors from the pre-trained features as well as the internal prior of the test image explicitly. Experimental results show that AdaFill outperforms other models on the various out-of-distribution test images. Furthermore, the model named ZeroFill, that are not pre-trained also sometimes outperforms the pre-trained models.

MyeongAh Cho, Taeoh Kim, Woo Jin Kim et al.

In contemporary society, surveillance anomaly detection, i.e., spotting anomalous events such as crimes or accidents in surveillance videos, is a critical task. As anomalies occur rarely, most training data consists of unlabeled videos without anomalous events, which makes the task challenging. Most existing methods use an autoencoder (AE) to learn to reconstruct normal videos; they then detect anomalies based on their failure to reconstruct the appearance of abnormal scenes. However, because anomalies are distinguished by appearance as well as motion, many previous approaches have explicitly separated appearance and motion information-for example, using a pre-trained optical flow model. This explicit separation restricts reciprocal representation capabilities between two types of information. In contrast, we propose an implicit two-path AE (ITAE), a structure in which two encoders implicitly model appearance and motion features, along with a single decoder that combines them to learn normal video patterns. For the complex distribution of normal scenes, we suggest normal density estimation of ITAE features through normalizing flow (NF)-based generative models to learn the tractable likelihoods and identify anomalies using out of distribution detection. NF models intensify ITAE performance by learning normality through implicitly learned features. Finally, we demonstrate the effectiveness of ITAE and its feature distribution modeling on six benchmarks, including databases that contain various anomalies in real-world scenarios.

Hyeongmin Lee, Taeoh Kim, Hanbin Son et al.

In Convolutional Neural Network (CNN) based image processing, most studies propose networks that are optimized to single-level (or single-objective); thus, they underperform on other levels and must be retrained for delivery of optimal performance. Using multiple models to cover multiple levels involves very high computational costs. To solve these problems, recent approaches train networks on two different levels and propose their own interpolation methods to enable arbitrary intermediate levels. However, many of them fail to generalize or have certain side effects in practical usage. In this paper, we define these frameworks as network tuning and interpolation and propose a novel module for continuous-level learning, called Filter Transition Network (FTN). This module is a structurally smoother module than existing ones. Therefore, the frameworks with FTN generalize well across various tasks and networks and cause fewer undesirable side effects. For stable learning of FTN, we additionally propose a method to initialize non-linear neural network layers with identity mappings. Extensive results for various image processing tasks indicate that the performance of FTN is comparable in multiple continuous levels, and is significantly smoother and lighter than that of other frameworks.

Hanbin Son, Taeoh Kim, Hyeongmin Lee et al.

To enhance image compression performance, recent deep neural network-based research can be divided into three categories: a learnable codec, a postprocessing network, and a compact representation network. The learnable codec has been designed for an end-to-end learning beyond the conventional compression modules. The postprocessing network increases the quality of decoded images using an example-based learning. The compact representation network is learned to reduce the capacity of an input image to reduce the bitrate while keeping the quality of the decoded image. However, these approaches are not compatible with the existing codecs or not optimal to increase the coding efficiency. Specifically, it is difficult to achieve optimal learning in the previous studies using the compact representation network, due to the inaccurate consideration of the codecs. In this paper, we propose a novel standard compatible image compression framework based on Auxiliary Codec Networks (ACNs). ACNs are designed to imitate image degradation operations of the existing codec, which delivers more accurate gradients to the compact representation network. Therefore, the compact representation and the postprocessing networks can be learned effectively and optimally. We demonstrate that our proposed framework based on JPEG and High Efficiency Video Coding (HEVC) standard substantially outperforms existing image compression algorithms in a standard compatible manner.

Sangjin Lee, Hyeongmin Lee, Taeoh Kim et al.

Video frame extrapolation is a task to predict future frames when the past frames are given. Unlike previous studies that usually have been focused on the design of modules or construction of networks, we propose a novel Extrapolative-Interpolative Cycle (EIC) loss using pre-trained frame interpolation module to improve extrapolation performance. Cycle-consistency loss has been used for stable prediction between two function spaces in many visual tasks. We formulate this cycle-consistency using two mapping functions; frame extrapolation and interpolation. Since it is easier to predict intermediate frames than to predict future frames in terms of the object occlusion and motion uncertainty, interpolation module can give guidance signal effectively for training the extrapolation function. EIC loss can be applied to any existing extrapolation algorithms and guarantee consistent prediction in the short future as well as long future frames. Experimental results show that simply adding EIC loss to the existing baseline increases extrapolation performance on both UCF101 and KITTI datasets.

Hyeongmin Lee, Taeoh Kim, Hanbin Son et al.

In Convolutional Neural Network (CNN) based image processing, most of the studies propose networks that are optimized for a single-level (or a single-objective); thus, they underperform on other levels and must be retrained for delivery of optimal performance. Using multiple models to cover multiple levels involves very high computational costs. To solve these problems, recent approaches train the networks on two different levels and propose their own interpolation methods to enable the arbitrary intermediate levels. However, many of them fail to adapt hard tasks or interpolate smoothly, or the others still require large memory and computational cost. In this paper, we propose a novel continuous-level learning framework using a Filter Transition Network (FTN) which is a non-linear module that easily adapt to new levels, and is regularized to prevent undesirable side-effects. Additionally, for stable learning of FTN, we newly propose a method to initialize non-linear CNNs with identity mappings. Furthermore, FTN is extremely lightweight module since it is a data-independent module, which means it is not affected by the spatial resolution of the inputs. Extensive results for various image processing tasks indicate that the performance of FTN is stable in terms of adaptation and interpolation, and comparable to that of the other heavy frameworks.

MyeongAh Cho, Taeoh Kim, Ig-Jae Kim et al.

Heterogeneous Face Recognition (HFR) is a task that matches faces across two different domains such as visible light (VIS), near-infrared (NIR), or the sketch domain. Due to the lack of databases, HFR methods usually exploit the pre-trained features on a large-scale visual database that contain general facial information. However, these pre-trained features cause performance degradation due to the texture discrepancy with the visual domain. With this motivation, we propose a graph-structured module called Relational Graph Module (RGM) that extracts global relational information in addition to general facial features. Because each identity's relational information between intra-facial parts is similar in any modality, the modeling relationship between features can help cross-domain matching. Through the RGM, relation propagation diminishes texture dependency without losing its advantages from the pre-trained features. Furthermore, the RGM captures global facial geometrics from locally correlated convolutional features to identify long-range relationships. In addition, we propose a Node Attention Unit (NAU) that performs node-wise recalibration to concentrate on the more informative nodes arising from relation-based propagation. Furthermore, we suggest a novel conditional-margin loss function (C-softmax) for the efficient projection learning of the embedding vector in HFR. The proposed method outperforms other state-of-the-art methods on five HFR databases. Furthermore, we demonstrate performance improvement on three backbones because our module can be plugged into any pre-trained face recognition backbone to overcome the limitations of a small HFR database.

Hyeongmin Lee, Taeoh Kim, Tae-young Chung et al.

Video frame interpolation is one of the most challenging tasks in video processing research. Recently, many studies based on deep learning have been suggested. Most of these methods focus on finding locations with useful information to estimate each output pixel using their own frame warping operations. However, many of them have Degrees of Freedom (DoF) limitations and fail to deal with the complex motions found in real world videos. To solve this problem, we propose a new warping module named Adaptive Collaboration of Flows (AdaCoF). Our method estimates both kernel weights and offset vectors for each target pixel to synthesize the output frame. AdaCoF is one of the most generalized warping modules compared to other approaches, and covers most of them as special cases of it. Therefore, it can deal with a significantly wide domain of complex motions. To further improve our framework and synthesize more realistic outputs, we introduce dual-frame adversarial loss which is applicable only to video frame interpolation tasks. The experimental results show that our method outperforms the state-of-the-art methods for both fixed training set environments and the Middlebury benchmark.