Hyunwoo J. Kim

Seongjun Yun, Seoyoon Kim, Junhyun Lee et al.

Graph Neural Networks (GNNs) have been widely applied to various fields for learning over graph-structured data. They have shown significant improvements over traditional heuristic methods in various tasks such as node classification and graph classification. However, since GNNs heavily rely on smoothed node features rather than graph structure, they often show poor performance than simple heuristic methods in link prediction where the structural information, e.g., overlapped neighborhoods, degrees, and shortest paths, is crucial. To address this limitation, we propose Neighborhood Overlap-aware Graph Neural Networks (Neo-GNNs) that learn useful structural features from an adjacency matrix and estimate overlapped neighborhoods for link prediction. Our Neo-GNNs generalize neighborhood overlap-based heuristic methods and handle overlapped multi-hop neighborhoods. Our extensive experiments on Open Graph Benchmark datasets (OGB) demonstrate that Neo-GNNs consistently achieve state-of-the-art performance in link prediction. Our code is publicly available at https://github.com/seongjunyun/Neo_GNNs.

Jinyoung Park, Sanghyeok Lee, Sihyeon Kim et al.

Transformers have shown superior performance on various computer vision tasks with their capabilities to capture long-range dependencies. Despite the success, it is challenging to directly apply Transformers on point clouds due to their quadratic cost in the number of points. In this paper, we present a Self-Positioning point-based Transformer (SPoTr), which is designed to capture both local and global shape contexts with reduced complexity. Specifically, this architecture consists of local self-attention and self-positioning point-based global cross-attention. The self-positioning points, adaptively located based on the input shape, consider both spatial and semantic information with disentangled attention to improve expressive power. With the self-positioning points, we propose a novel global cross-attention mechanism for point clouds, which improves the scalability of global self-attention by allowing the attention module to compute attention weights with only a small set of self-positioning points. Experiments show the effectiveness of SPoTr on three point cloud tasks such as shape classification, part segmentation, and scene segmentation. In particular, our proposed model achieves an accuracy gain of 2.6% over the previous best models on shape classification with ScanObjectNN. We also provide qualitative analyses to demonstrate the interpretability of self-positioning points. The code of SPoTr is available at https://github.com/mlvlab/SPoTr.

Hyeong Kyu Choi, Joonmyung Choi, Hyunwoo J. Kim

Mixup is a commonly adopted data augmentation technique for image classification. Recent advances in mixup methods primarily focus on mixing based on saliency. However, many saliency detectors require intense computation and are especially burdensome for parameter-heavy transformer models. To this end, we propose TokenMixup, an efficient attention-guided token-level data augmentation method that aims to maximize the saliency of a mixed set of tokens. TokenMixup provides x15 faster saliency-aware data augmentation compared to gradient-based methods. Moreover, we introduce a variant of TokenMixup which mixes tokens within a single instance, thereby enabling multi-scale feature augmentation. Experiments show that our methods significantly improve the baseline models' performance on CIFAR and ImageNet-1K, while being more efficient than previous methods. We also reach state-of-the-art performance on CIFAR-100 among from-scratch transformer models. Code is available at https://github.com/mlvlab/TokenMixup.

Dongjun Lee, Seokwon Song, Jihee Suh et al.

In recent years, prompt tuning has proven effective in adapting pre-trained vision-language models to downstream tasks. These methods aim to adapt the pre-trained models by introducing learnable prompts while keeping pre-trained weights frozen. However, learnable prompts can affect the internal representation within the self-attention module, which may negatively impact performance variance and generalization, especially in data-deficient settings. To address these issues, we propose a novel approach, Read-only Prompt Optimization (RPO). RPO leverages masked attention to prevent the internal representation shift in the pre-trained model. Further, to facilitate the optimization of RPO, the read-only prompts are initialized based on special tokens of the pre-trained model. Our extensive experiments demonstrate that RPO outperforms CLIP and CoCoOp in base-to-new generalization and domain generalization while displaying better robustness. Also, the proposed method achieves better generalization on extremely data-deficient settings, while improving parameter efficiency and computational overhead. Code is available at https://github.com/mlvlab/RPO.

Dohwan Ko, Joonmyung Choi, Hyeong Kyu Choi et al.

Foundation models have shown outstanding performance and generalization capabilities across domains. Since most studies on foundation models mainly focus on the pretraining phase, a naive strategy to minimize a single task-specific loss is adopted for fine-tuning. However, such fine-tuning methods do not fully leverage other losses that are potentially beneficial for the target task. Therefore, we propose MEta Loss TRansformer (MELTR), a plug-in module that automatically and non-linearly combines various loss functions to aid learning the target task via auxiliary learning. We formulate the auxiliary learning as a bi-level optimization problem and present an efficient optimization algorithm based on Approximate Implicit Differentiation (AID). For evaluation, we apply our framework to various video foundation models (UniVL, Violet and All-in-one), and show significant performance gain on all four downstream tasks: text-to-video retrieval, video question answering, video captioning, and multi-modal sentiment analysis. Our qualitative analyses demonstrate that MELTR adequately `transforms' individual loss functions and `melts' them into an effective unified loss. Code is available at https://github.com/mlvlab/MELTR.

Jihwan Park, SeungJun Lee, Hwan Heo et al.

Human-Object Interaction detection is a holistic visual recognition task that entails object detection as well as interaction classification. Previous works of HOI detection has been addressed by the various compositions of subset predictions, e.g., Image -> HO -> I, Image -> HI -> O. Recently, transformer based architecture for HOI has emerged, which directly predicts the HOI triplets in an end-to-end fashion (Image -> HOI). Motivated by various inference paths for HOI detection, we propose cross-path consistency learning (CPC), which is a novel end-to-end learning strategy to improve HOI detection for transformers by leveraging augmented decoding paths. CPC learning enforces all the possible predictions from permuted inference sequences to be consistent. This simple scheme makes the model learn consistent representations, thereby improving generalization without increasing model capacity. Our experiments demonstrate the effectiveness of our method, and we achieved significant improvement on V-COCO and HICO-DET compared to the baseline models. Our code is available at https://github.com/mlvlab/CPChoi.

Jinyoung Park, Hyeong Kyu Choi, Juyeon Ko et al.

Question Answering (QA) is a task that entails reasoning over natural language contexts, and many relevant works augment language models (LMs) with graph neural networks (GNNs) to encode the Knowledge Graph (KG) information. However, most existing GNN-based modules for QA do not take advantage of rich relational information of KGs and depend on limited information interaction between the LM and the KG. To address these issues, we propose Question Answering Transformer (QAT), which is designed to jointly reason over language and graphs with respect to entity relations in a unified manner. Specifically, QAT constructs Meta-Path tokens, which learn relation-centric embeddings based on diverse structural and semantic relations. Then, our Relation-Aware Self-Attention module comprehensively integrates different modalities via the Cross-Modal Relative Position Bias, which guides information exchange between relevant entites of different modalities. We validate the effectiveness of QAT on commonsense question answering datasets like CommonsenseQA and OpenBookQA, and on a medical question answering dataset, MedQA-USMLE. On all the datasets, our method achieves state-of-the-art performance. Our code is available at http://github.com/mlvlab/QAT.

Minkyu Jeon, Hyeonjin Park, Hyunwoo J. Kim et al.

The application of modern machine learning to retinal image analyses offers valuable insights into a broad range of human health conditions beyond ophthalmic diseases. Additionally, data sharing is key to fully realizing the potential of machine learning models by providing a rich and diverse collection of training data. However, the personally-identifying nature of retinal images, encompassing the unique vascular structure of each individual, often prevents this data from being shared openly. While prior works have explored image de-identification strategies based on synthetic averaging of images in other domains (e.g. facial images), existing techniques face difficulty in preserving both privacy and clinical utility in retinal images, as we demonstrate in our work. We therefore introduce k-SALSA, a generative adversarial network (GAN)-based framework for synthesizing retinal fundus images that summarize a given private dataset while satisfying the privacy notion of k-anonymity. k-SALSA brings together state-of-the-art techniques for training and inverting GANs to achieve practical performance on retinal images. Furthermore, k-SALSA leverages a new technique, called local style alignment, to generate a synthetic average that maximizes the retention of fine-grain visual patterns in the source images, thus improving the clinical utility of the generated images. On two benchmark datasets of diabetic retinopathy (EyePACS and APTOS), we demonstrate our improvement upon existing methods with respect to image fidelity, classification performance, and mitigation of membership inference attacks. Our work represents a step toward broader sharing of retinal images for scientific collaboration. Code is available at https://github.com/hcholab/k-salsa.

Minseok Joo, Dogyun Park, Taehoon Lee et al.

Maintaining long-term geometric consistency remains challenging for long-horizon autoregressive video generation. Memory-augmented generative models address this by retrieving historical frames, but their effectiveness depends on two key design choices: what 3D-geometric evidence should represent past observations, and how memory frames should be selected from this evidence. Existing methods often rely on camera poses or field-of-view overlap, which are lightweight but too coarse to reason about pixel-wise visibility, or use explicit 3D reconstruction, which provides fine-grained evidence but is costly to maintain over long rollouts. We propose Coverage-Maximizing Retrieval-Augmented Generation (COVRAG), a depth-based memory retrieval framework that uses pretrained 3D priors to construct a target-view coverage map as lightweight 3D memory evidence. For frame selection, COVRAG maximizes residual coverage gain, iteratively retrieving frames that explain target-view regions not covered by the current context or previously selected memories. To improve scalability in long-video generation, we introduce sliding-window depth caching for efficient geometry estimation. Experiments on RealEstate10K and DL3DV10K show that COVRAG improves long-horizon geometric consistency while maintaining low latency compared to baselines.

Eulrang Cho, Jooyeon Kim, Hyunwoo J. Kim

Pre-trained vision-language models (VLMs) have shown impressive performance on various downstream tasks by utilizing knowledge learned from large data. In general, the performance of VLMs on target tasks can be further improved by prompt tuning, which adds context to the input image or text. By leveraging data from target tasks, various prompt-tuning methods have been studied in the literature. A key to prompt tuning is the feature space alignment between two modalities via learnable vectors with model parameters fixed. We observed that the alignment becomes more effective when embeddings of each modality are `well-arranged' in the latent space. Inspired by this observation, we proposed distribution-aware prompt tuning (DAPT) for vision-language models, which is simple yet effective. Specifically, the prompts are learned by maximizing inter-dispersion, the distance between classes, as well as minimizing the intra-dispersion measured by the distance between embeddings from the same class. Our extensive experiments on 11 benchmark datasets demonstrate that our method significantly improves generalizability. The code is available at https://github.com/mlvlab/DAPT.

Sihyeon Kim, Minseok Joo, Jaewon Lee et al.

Learning implicit templates as neural fields has recently shown impressive performance in unsupervised shape correspondence. Despite the success, we observe current approaches, which solely rely on geometric information, often learn suboptimal deformation across generic object shapes, which have high structural variability. In this paper, we highlight the importance of part deformation consistency and propose a semantic-aware implicit template learning framework to enable semantically plausible deformation. By leveraging semantic prior from a self-supervised feature extractor, we suggest local conditioning with novel semantic-aware deformation code and deformation consistency regularizations regarding part deformation, global deformation, and global scaling. Our extensive experiments demonstrate the superiority of the proposed method over baselines in various tasks: keypoint transfer, part label transfer, and texture transfer. More interestingly, our framework shows a larger performance gain under more challenging settings. We also provide qualitative analyses to validate the effectiveness of semantic-aware deformation. The code is available at https://github.com/mlvlab/PDC.

Hyeong Kyu Choi, Seunghun Lee, Jaewon Chu et al.

Multi-hop Knowledge Graph Question Answering (KGQA) is a task that involves retrieving nodes from a knowledge graph (KG) to answer natural language questions. Recent GNN-based approaches formulate this task as a KG path searching problem, where messages are sequentially propagated from the seed node towards the answer nodes. However, these messages are past-oriented, and they do not consider the full KG context. To make matters worse, KG nodes often represent proper noun entities and are sometimes encrypted, being uninformative in selecting between paths. To address these problems, we propose Neural Tree Search (NuTrea), a tree search-based GNN model that incorporates the broader KG context. Our model adopts a message-passing scheme that probes the unreached subtree regions to boost the past-oriented embeddings. In addition, we introduce the Relation Frequency-Inverse Entity Frequency (RF-IEF) node embedding that considers the global KG context to better characterize ambiguous KG nodes. The general effectiveness of our approach is demonstrated through experiments on three major multi-hop KGQA benchmark datasets, and our extensive analyses further validate its expressiveness and robustness. Overall, NuTrea provides a powerful means to query the KG with complex natural language questions. Code is available at https://github.com/mlvlab/NuTrea.

Dohwan Ko, Ji Soo Lee, Wooyoung Kang et al.

Large Language Models (LLMs) have shown remarkable performances on a wide range of natural language understanding and generation tasks. We observe that the LLMs provide effective priors in exploiting $\textit{linguistic shortcuts}$ for temporal and causal reasoning in Video Question Answering (VideoQA). However, such priors often cause suboptimal results on VideoQA by leading the model to over-rely on questions, $\textit{i.e.}$, $\textit{linguistic bias}$, while ignoring visual content. This is also known as `ungrounded guesses' or `hallucinations'. To address this problem while leveraging LLMs' prior on VideoQA, we propose a novel framework, Flipped-VQA, encouraging the model to predict all the combinations of $\langle$V, Q, A$\rangle$ triplet by flipping the source pair and the target label to understand their complex relationships, $\textit{i.e.}$, predict A, Q, and V given a VQ, VA, and QA pairs, respectively. In this paper, we develop LLaMA-VQA by applying Flipped-VQA to LLaMA, and it outperforms both LLMs-based and non-LLMs-based models on five challenging VideoQA benchmarks. Furthermore, our Flipped-VQA is a general framework that is applicable to various LLMs (OPT and GPT-J) and consistently improves their performances. We empirically demonstrate that Flipped-VQA not only enhances the exploitation of linguistic shortcuts but also mitigates the linguistic bias, which causes incorrect answers over-relying on the question. Code is available at https://github.com/mlvlab/Flipped-VQA.

Byeongkeun Ahn, Chiyoon Kim, Youngjoon Hong et al.

Normalizing flows model probability distributions by learning invertible transformations that transfer a simple distribution into complex distributions. Since the architecture of ResNet-based normalizing flows is more flexible than that of coupling-based models, ResNet-based normalizing flows have been widely studied in recent years. Despite their architectural flexibility, it is well-known that the current ResNet-based models suffer from constrained Lipschitz constants. In this paper, we propose the monotone formulation to overcome the issue of the Lipschitz constants using monotone operators and provide an in-depth theoretical analysis. Furthermore, we construct an activation function called Concatenated Pila (CPila) to improve gradient flow. The resulting model, Monotone Flows, exhibits an excellent performance on multiple density estimation benchmarks (MNIST, CIFAR-10, ImageNet32, ImageNet64). Code is available at https://github.com/mlvlab/MonotoneFlows.

Hyeonjin Park, Seunghun Lee, Sihyeon Kim et al.

Graph Neural Networks (GNNs) often suffer from weak-generalization due to sparsely labeled data despite their promising results on various graph-based tasks. Data augmentation is a prevalent remedy to improve the generalization ability of models in many domains. However, due to the non-Euclidean nature of data space and the dependencies between samples, designing effective augmentation on graphs is challenging. In this paper, we propose a novel framework Metropolis-Hastings Data Augmentation (MH-Aug) that draws augmented graphs from an explicit target distribution for semi-supervised learning. MH-Aug produces a sequence of augmented graphs from the target distribution enables flexible control of the strength and diversity of augmentation. Since the direct sampling from the complex target distribution is challenging, we adopt the Metropolis-Hastings algorithm to obtain the augmented samples. We also propose a simple and effective semi-supervised learning strategy with generated samples from MH-Aug. Our extensive experiments demonstrate that MH-Aug can generate a sequence of samples according to the target distribution to significantly improve the performance of GNNs.

Sojin Lee, Dogyun Park, Inho Kong et al.

Recent studies on inverse problems have proposed posterior samplers that leverage the pre-trained diffusion models as powerful priors. These attempts have paved the way for using diffusion models in a wide range of inverse problems. However, the existing methods entail computationally demanding iterative sampling procedures and optimize a separate solution for each measurement, which leads to limited scalability and lack of generalization capability across unseen samples. To address these limitations, we propose a novel approach, Diffusion prior-based Amortized Variational Inference (DAVI) that solves inverse problems with a diffusion prior from an amortized variational inference perspective. Specifically, instead of separate measurement-wise optimization, our amortized inference learns a function that directly maps measurements to the implicit posterior distributions of corresponding clean data, enabling a single-step posterior sampling even for unseen measurements. Extensive experiments on image restoration tasks, e.g., Gaussian deblur, 4$\times$ super-resolution, and box inpainting with two benchmark datasets, demonstrate our approach's superior performance over strong baselines. Code is available at https://github.com/mlvlab/DAVI.

Kyujin Lee, Injae Kim, Jihwan Park et al.

Vision-Language-Action (VLA) models have emerged as a promising framework that unifies perception, reasoning, and control for robot manipulation by adapting pretrained vision-language models (VLMs) to action prediction. However, VLM-derived representations are often insensitive to subtle visual distinctions required for low-level control, causing state aliasing between visually similar states that require substantially different actions. Prior VLA studies improve visual understanding by generating visual or reasoning outputs, such as future frames, 2D grounding points or traces, or intermediate spatial reasoning steps, but these objectives typically shape the vision encoder only indirectly through end-to-end prediction and do not explicitly analyze state aliasing in the learned visual feature space. To mitigate state aliasing, we introduce inverse dynamics learning as an auxiliary objective that directly supervises the VLA vision encoder. By predicting the action between current and future observations, our objective encourages the encoder to capture fine-grained visual distinctions that determine low-level actions. We further use pseudo-reversed supervision to expose the encoder to a broader range of action directions and improve generalization under limited robot demonstrations. Our method applies to diverse VLA baselines, uses only standard observation-action pairs without additional annotations, and preserves the original inference pipeline at test time. Experiments on CALVIN ABC-D and SimplerEnv show consistent gains across diverse VLA baselines. Frozen-encoder probing and state-feature alignment analyses further show that our method learns state-discriminative visual representations that reduce state aliasing and better align with robot state changes.

Dohwan Ko, Ji Soo Lee, Miso Choi et al.

Video Question Answering (VideoQA) is a challenging task that entails complex multi-modal reasoning. In contrast to multiple-choice VideoQA which aims to predict the answer given several options, the goal of open-ended VideoQA is to answer questions without restricting candidate answers. However, the majority of previous VideoQA models formulate open-ended VideoQA as a classification task to classify the video-question pairs into a fixed answer set, i.e., closed-vocabulary, which contains only frequent answers (e.g., top-1000 answers). This leads the model to be biased toward only frequent answers and fail to generalize on out-of-vocabulary answers. We hence propose a new benchmark, Open-vocabulary Video Question Answering (OVQA), to measure the generalizability of VideoQA models by considering rare and unseen answers. In addition, in order to improve the model's generalization power, we introduce a novel GNN-based soft verbalizer that enhances the prediction on rare and unseen answers by aggregating the information from their similar words. For evaluation, we introduce new baselines by modifying the existing (closed-vocabulary) open-ended VideoQA models and improve their performances by further taking into account rare and unseen answers. Our ablation studies and qualitative analyses demonstrate that our GNN-based soft verbalizer further improves the model performance, especially on rare and unseen answers. We hope that our benchmark OVQA can serve as a guide for evaluating the generalizability of VideoQA models and inspire future research. Code is available at https://github.com/mlvlab/OVQA.

Hwan Heo, Taekyung Kim, Jiyoung Lee et al.

Multi-resolution hash encoding has recently been proposed to reduce the computational cost of neural renderings, such as NeRF. This method requires accurate camera poses for the neural renderings of given scenes. However, contrary to previous methods jointly optimizing camera poses and 3D scenes, the naive gradient-based camera pose refinement method using multi-resolution hash encoding severely deteriorates performance. We propose a joint optimization algorithm to calibrate the camera pose and learn a geometric representation using efficient multi-resolution hash encoding. Showing that the oscillating gradient flows of hash encoding interfere with the registration of camera poses, our method addresses the issue by utilizing smooth interpolation weighting to stabilize the gradient oscillation for the ray samplings across hash grids. Moreover, the curriculum training procedure helps to learn the level-wise hash encoding, further increasing the pose refinement. Experiments on the novel-view synthesis datasets validate that our learning frameworks achieve state-of-the-art performance and rapid convergence of neural rendering, even when initial camera poses are unknown.

Jinyoung Park, Seongjun Yun, Hyeonjin Park et al.

Transformer-based models have recently shown success in representation learning on graph-structured data beyond natural language processing and computer vision. However, the success is limited to small-scale graphs due to the drawbacks of full dot-product attention on graphs such as the quadratic complexity with respect to the number of nodes and message aggregation from enormous irrelevant nodes. To address these issues, we propose Deformable Graph Transformer (DGT) that performs sparse attention via dynamically sampled relevant nodes for efficiently handling large-scale graphs with a linear complexity in the number of nodes. Specifically, our framework first constructs multiple node sequences with various criteria to consider both structural and semantic proximity. Then, combining with our learnable Katz Positional Encodings, the sparse attention is applied to the node sequences for learning node representations with a significantly reduced computational cost. Extensive experiments demonstrate that our DGT achieves state-of-the-art performance on 7 graph benchmark datasets with 2.5 - 449 times less computational cost compared to transformer-based graph models with full attention.

Sanghyeok Lee, Minkyu Jeon, Injae Kim et al.

Data augmentation is key to improving the generalization ability of deep learning models. Mixup is a simple and widely-used data augmentation technique that has proven effective in alleviating the problems of overfitting and data scarcity. Also, recent studies of saliency-aware Mixup in the image domain show that preserving discriminative parts is beneficial to improving the generalization performance. However, these Mixup-based data augmentations are underexplored in 3D vision, especially in point clouds. In this paper, we propose SageMix, a saliency-guided Mixup for point clouds to preserve salient local structures. Specifically, we extract salient regions from two point clouds and smoothly combine them into one continuous shape. With a simple sequential sampling by re-weighted saliency scores, SageMix preserves the local structure of salient regions. Extensive experiments demonstrate that the proposed method consistently outperforms existing Mixup methods in various benchmark point cloud datasets. With PointNet++, our method achieves an accuracy gain of 2.6% and 4.0% over standard training in 3D Warehouse dataset (MN40) and ScanObjectNN, respectively. In addition to generalization performance, SageMix improves robustness and uncertainty calibration. Moreover, when adopting our method to various tasks including part segmentation and standard 2D image classification, our method achieves competitive performance.

Juhan Cha, Minseok Joo, Jihwan Park et al.

Recent advancements in 3D object detection have benefited from multi-modal information from the multi-view cameras and LiDAR sensors. However, the inherent disparities between the modalities pose substantial challenges. We observe that existing multi-modal 3D object detection methods heavily rely on the LiDAR sensor, treating the camera as an auxiliary modality for augmenting semantic details. This often leads to not only underutilization of camera data but also significant performance degradation in scenarios where LiDAR data is unavailable. Additionally, existing fusion methods overlook the detrimental impact of sensor noise induced by environmental changes, on detection performance. In this paper, we propose MEFormer to address the LiDAR over-reliance problem by harnessing critical information for 3D object detection from every available modality while concurrently safeguarding against corrupted signals during the fusion process. Specifically, we introduce Modality Agnostic Decoding (MOAD) that extracts geometric and semantic features with a shared transformer decoder regardless of input modalities and provides promising improvement with a single modality as well as multi-modality. Additionally, our Proximity-based Modality Ensemble (PME) module adaptively utilizes the strengths of each modality depending on the environment while mitigating the effects of a noisy sensor. Our MEFormer achieves state-of-the-art performance of 73.9% NDS and 71.5% mAP in the nuScenes validation set. Extensive analyses validate that our MEFormer improves robustness against challenging conditions such as sensor malfunctions or environmental changes. The source code is available at https://github.com/hanchaa/MEFormer

Jaewon Lee, Injae Kim, Hwan Heo et al.

We present a learning framework for reconstructing neural scene representations from a small number of unconstrained tourist photos. Since each image contains transient occluders, decomposing the static and transient components is necessary to construct radiance fields with such in-the-wild photographs where existing methods require a lot of training data. We introduce SF-NeRF, aiming to disentangle those two components with only a few images given, which exploits semantic information without any supervision. The proposed method contains an occlusion filtering module that predicts the transient color and its opacity for each pixel, which enables the NeRF model to solely learn the static scene representation. This filtering module learns the transient phenomena guided by pixel-wise semantic features obtained by a trainable image encoder that can be trained across multiple scenes to learn the prior of transient objects. Furthermore, we present two techniques to prevent ambiguous decomposition and noisy results of the filtering module. We demonstrate that our method outperforms state-of-the-art novel view synthesis methods on Phototourism dataset in a few-shot setting.

Jihwan Park, Chanhyeong Yang, Jinyoung Park et al.

Weakly-supervised Human-Object Interaction (HOI) detection is essential for scalable scene understanding, as it learns interactions from only image-level annotations. Due to the lack of localization signals, prior works typically rely on an external object detector to generate candidate pairs and then infer their interactions through pairwise reasoning. However, this framework often struggles to scale due to the substantial computational cost incurred by enumerating numerous instance pairs. In addition, it suffers from false positives arising from non-interactive combinations, which hinder accurate instance-level HOI reasoning. To address these issues, we introduce Relational Grounding Transformer (RegFormer), a versatile interaction recognition module for efficient and accurate HOI reasoning. Under image-level supervision, RegFormer leverages spatially grounded signals as guidance for the reasoning process and promotes locality-aware interaction learning. By learning localized interaction cues, our module distinguishes humans, objects, and their interactions, enabling direct transfer from image-level interaction reasoning to precise and efficient instance-level reasoning without additional training. Our extensive experiments and analyses demonstrate that RegFormer effectively learns spatial cues for instance-level interaction reasoning, operates with high efficiency, and even achieves performance comparable to fully supervised models. Our code is available at https://github.com/mlvlab/RegFormer.

Inho Kong, Sojin Lee, Youngjoon Hong et al.

Classifier-Free Guidance (CFG) has established the foundation for guidance mechanisms in diffusion models, showing that well-designed guidance proxies significantly improve conditional generation and sample quality. Autoguidance (AG) has extended this idea, but it relies on an auxiliary network and leaves solver-induced errors unaddressed. In stiff regions, the ODE trajectory changes sharply, where local truncation error (LTE) becomes a critical factor that deteriorates sample quality. Our key observation is that these errors align with the dominant eigenvector, motivating us to leverage the solver-induced error as a guidance signal. We propose Embedded Runge-Kutta Guidance (ERK-Guid), which exploits detected stiffness to reduce LTE and stabilize sampling. We theoretically and empirically analyze stiffness and eigenvector estimators with solver errors to motivate the design of ERK-Guid. Our experiments on both synthetic datasets and the popular benchmark dataset, ImageNet, demonstrate that ERK-Guid consistently outperforms state-of-the-art methods. Code is available at https://github.com/mlvlab/ERK-Guid.

Injae Kim, Jongha Kim, Joonmyung Choi et al.

Interpretability is a crucial factor in building reliable models for various medical applications. Concept Bottleneck Models (CBMs) enable interpretable image classification by utilizing human-understandable concepts as intermediate targets. Unlike conventional methods that require extensive human labor to construct the concept set, recent works leveraging Large Language Models (LLMs) for generating concepts made automatic concept generation possible. However, those methods do not consider whether a concept is visually relevant or not, which is an important factor in computing meaningful concept scores. Therefore, we propose a visual activation score that measures whether the concept contains visual cues or not, which can be easily computed with unlabeled image data. Computed visual activation scores are then used to filter out the less visible concepts, thus resulting in a final concept set with visually meaningful concepts. Our experimental results show that adopting the proposed visual activation score for concept filtering consistently boosts performance compared to the baseline. Moreover, qualitative analyses also validate that visually relevant concepts are successfully selected with the visual activation score.

Hwan Heo, Youngjin Oh, Jaewon Lee et al.

Recent studies have proven that DNNs, unlike human vision, tend to exploit texture information rather than shape. Such texture bias is one of the factors for the poor generalization performance of DNNs. We observe that the texture bias negatively affects not only in-domain generalization but also out-of-distribution generalization, i.e., Domain Generalization. Motivated by the observation, we propose a new framework to reduce the texture bias of a model by a novel optimization-based data augmentation, dubbed Stylized Dream. Our framework utilizes adaptive instance normalization (AdaIN) to augment the style of an original image yet preserve the content. We then adopt a regularization loss to predict consistent outputs between Stylized Dream and original images, which encourages the model to learn shape-based representations. Extensive experiments show that the proposed method achieves state-of-the-art performance in out-of-distribution settings on public benchmark datasets: PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet.

Seunghun Lee, Jaewon Chu, Sihyeon Kim et al.

Bayesian optimization is a powerful method for optimizing black-box functions with limited function evaluations. Recent works have shown that optimization in a latent space through deep generative models such as variational autoencoders leads to effective and efficient Bayesian optimization for structured or discrete data. However, as the optimization does not take place in the input space, it leads to an inherent gap that results in potentially suboptimal solutions. To alleviate the discrepancy, we propose Correlated latent space Bayesian Optimization (CoBO), which focuses on learning correlated latent spaces characterized by a strong correlation between the distances in the latent space and the distances within the objective function. Specifically, our method introduces Lipschitz regularization, loss weighting, and trust region recoordination to minimize the inherent gap around the promising areas. We demonstrate the effectiveness of our approach on several optimization tasks in discrete data, such as molecule design and arithmetic expression fitting, and achieve high performance within a small budget.

Joonmyung Choi, Sanghyeok Lee, Jaewon Chu et al.

Video Transformers have become the prevalent solution for various video downstream tasks with superior expressive power and flexibility. However, these video transformers suffer from heavy computational costs induced by the massive number of tokens across the entire video frames, which has been the major barrier to training the model. Further, the patches irrelevant to the main contents, e.g., backgrounds, degrade the generalization performance of models. To tackle these issues, we propose training free token merging for lightweight video Transformer (vid-TLDR) that aims to enhance the efficiency of video Transformers by merging the background tokens without additional training. For vid-TLDR, we introduce a novel approach to capture the salient regions in videos only with the attention map. Further, we introduce the saliency-aware token merging strategy by dropping the background tokens and sharpening the object scores. Our experiments show that vid-TLDR significantly mitigates the computational complexity of video Transformers while achieving competitive performance compared to the base model without vid-TLDR. Code is available at https://github.com/mlvlab/vid-TLDR.

Jinyoung Park, Juyeon Ko, Hyunwoo J. Kim

Pre-trained vision-language models have shown impressive success on various computer vision tasks with their zero-shot generalizability. Recently, prompt learning approaches have been explored to efficiently and effectively adapt the vision-language models to a variety of downstream tasks. However, most existing prompt learning methods suffer from task overfitting since the general knowledge of the pre-trained vision language models is forgotten while the prompts are finetuned on a small data set from a specific target task. To address this issue, we propose a Prompt Meta-Regularization (ProMetaR) to improve the generalizability of prompt learning for vision-language models. Specifically, ProMetaR meta-learns both the regularizer and the soft prompts to harness the task-specific knowledge from the downstream tasks and task-agnostic general knowledge from the vision-language models. Further, ProMetaR augments the task to generate multiple virtual tasks to alleviate the meta-overfitting. In addition, we provide the analysis to comprehend how ProMetaR improves the generalizability of prompt tuning in the perspective of the gradient alignment. Our extensive experiments demonstrate that our ProMetaR improves the generalizability of conventional prompt learning methods under base-to-base/base-to-new and domain generalization settings. The code of ProMetaR is available at https://github.com/mlvlab/ProMetaR.

Jinyoung Park, Ameen Patel, Omar Zia Khan et al.

Chain-of-Thought (CoT) prompting along with sub-question generation and answering has enhanced multi-step reasoning capabilities of Large Language Models (LLMs). However, prompting the LLMs to directly generate sub-questions is suboptimal since they sometimes generate redundant or irrelevant questions. To deal with them, we propose a GE-Reasoning method, which directs LLMs to generate proper sub-questions and corresponding answers. Concretely, given an input question, we first prompt the LLM to generate knowledge triplets, forming a graph representation of the question. Unlike conventional knowledge triplets, our approach allows variables as head or tail entities, effectively representing a question as knowledge triplets. Second, for each triplet, the LLM generates a corresponding sub-question and answer along with using knowledge retrieval. If the prediction confidence exceeds a threshold, the sub-question and prediction are incorporated into the prompt for subsequent processing. This approach encourages that sub-questions are grounded in the extracted knowledge triplets, reducing redundancy and irrelevance. Our experiments demonstrate that our approach outperforms previous CoT prompting methods and their variants on multi-hop question answering benchmark datasets.

Sanghyeok Lee, Joonmyung Choi, Hyunwoo J. Kim

Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion (MCTF), that gradually fuses the tokens based on multi-criteria (e.g., similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens. By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about 44% while improving the performance (+0.5%, and +0.3%) over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least 31% speedup without performance degradation. Code is available at https://github.com/mlvlab/MCTF.

Jooyeon Kim, Eulrang Cho, Sehyung Kim et al.

Open-vocabulary object detection (OVD) has been studied with Vision-Language Models (VLMs) to detect novel objects beyond the pre-trained categories. Previous approaches improve the generalization ability to expand the knowledge of the detector, using 'positive' pseudo-labels with additional 'class' names, e.g., sock, iPod, and alligator. To extend the previous methods in two aspects, we propose Retrieval-Augmented Losses and visual Features (RALF). Our method retrieves related 'negative' classes and augments loss functions. Also, visual features are augmented with 'verbalized concepts' of classes, e.g., worn on the feet, handheld music player, and sharp teeth. Specifically, RALF consists of two modules: Retrieval Augmented Losses (RAL) and Retrieval-Augmented visual Features (RAF). RAL constitutes two losses reflecting the semantic similarity with negative vocabularies. In addition, RAF augments visual features with the verbalized concepts from a large language model (LLM). Our experiments demonstrate the effectiveness of RALF on COCO and LVIS benchmark datasets. We achieve improvement up to 3.4 box AP$_{50}^{\text{N}}$ on novel categories of the COCO dataset and 3.6 mask AP$_{\text{r}}$ gains on the LVIS dataset. Code is available at https://github.com/mlvlab/RALF .

Dohwan Ko, Jinyoung Park, Seoung Choi et al.

Mixture-of-Experts (MoE) has emerged as an effective approach to reduce the computational overhead of Transformer architectures by sparsely activating a subset of parameters for each token while preserving high model capacity. This paradigm has recently been extended to Vision-Language Models (VLMs), enabling scalable multi-modal understanding with reduced computational cost. However, the widely adopted deterministic top-K routing mechanism may overlook more optimal expert combinations and lead to expert overfitting. To address this limitation and improve the diversity of expert selection, we propose MoE-GRPO, a reinforcement learning (RL)-based framework for optimizing expert routing in MoE-based VLMs. Specifically, we formulate expert selection as a sequential decision-making problem and optimize it using Group Relative Policy Optimization (GRPO), allowing the model to learn adaptive expert routing policies through exploration and reward-based feedback. Furthermore, we introduce a modality-aware router guidance that enhances training stability and efficiency by discouraging the router from exploring experts that are infrequently activated for a given modality. Extensive experiments on multi-modal image and video benchmarks show that MoE-GRPO consistently outperforms standard top-K routing and its variants by promoting more diverse expert selection, thereby mitigating expert overfitting and enabling a task-level expert specialization.

Joonmyung Choi, Sanghyeok Lee, Jongha Kim et al.

Recent advances in vision-language models have demonstrated remarkable performance across diverse multi-modal tasks, including document question answering that leverages structured visual cues from text, tables, and figures. However, unlike natural images, document images contain large backgrounds and only sparse supporting evidence, leading to the inefficient consumption of substantial computational resources, especially for long documents. We observe that existing token-reduction methods for natural images and videos fall short in utilizing the structural sparsity unique to documents. To address this, we propose DocPrune, a training-free and progressive document token pruning framework designed for efficient long-document understanding. The proposed method preserves only the essential tokens for the task while removing unnecessary ones, such as background or question-irrelevant tokens. Moreover, it automatically selects the appropriate layers to initiate token pruning based on the model's level of comprehension. Our experiments on the M3DocRAG show that DocPrune improves throughput by 3.0x and 3.3x in the encoder and decoder, respectively, while boosting the F1 score by +1.0, achieving both higher accuracy and efficiency without any additional training.

Jinyoung Park, Minseong Bae, Dohwan Ko et al.

Large Language Models (LLMs) have demonstrated remarkable generalization and instruction-following capabilities with instruction tuning. The advancements in LLMs and instruction tuning have led to the development of Large Vision-Language Models (LVLMs). However, the competency of the LLMs and instruction tuning have been less explored in the molecular domain. Thus, we propose LLaMo: Large Language Model-based Molecular graph assistant, which is an end-to-end trained large molecular graph-language model. To bridge the discrepancy between the language and graph modalities, we present the multi-level graph projector that transforms graph representations into graph tokens by abstracting the output representations of each GNN layer and motif representations with the cross-attention mechanism. We also introduce machine-generated molecular graph instruction data to instruction-tune the large molecular graph-language model for general-purpose molecule and language understanding. Our extensive experiments demonstrate that LLaMo shows the best performance on diverse tasks, such as molecular description generation, property prediction, and IUPAC name prediction. The code of LLaMo is available at https://github.com/mlvlab/LLaMo.

Injae Kim, Minhyuk Choi, Hyunwoo J. Kim

Neural Radiance Field (NeRF) has enabled novel view synthesis with high fidelity given images and camera poses. Subsequent works even succeeded in eliminating the necessity of pose priors by jointly optimizing NeRF and camera pose. However, these works are limited to relatively simple settings such as photometrically consistent and occluder-free image collections or a sequence of images from a video. So they have difficulty handling unconstrained images with varying illumination and transient occluders. In this paper, we propose $\textbf{UP-NeRF}$ ($\textbf{U}$nconstrained $\textbf{P}$ose-prior-free $\textbf{Ne}$ural $\textbf{R}$adiance $\textbf{F}$ields) to optimize NeRF with unconstrained image collections without camera pose prior. We tackle these challenges with surrogate tasks that optimize color-insensitive feature fields and a separate module for transient occluders to block their influence on pose estimation. In addition, we introduce a candidate head to enable more robust pose estimation and transient-aware depth supervision to minimize the effect of incorrect prior. Our experiments verify the superior performance of our method compared to the baselines including BARF and its variants in a challenging internet photo collection, $\textit{Phototourism}$ dataset.

Sanghyeok Lee, Joonmyung Choi, Hyunwoo J. Kim

For the deployment of neural networks in resource-constrained environments, prior works have built lightweight architectures with convolution and attention for capturing local and global dependencies, respectively. Recently, the state space model (SSM) has emerged as an effective operation for global interaction with its favorable linear computational cost in the number of tokens. To harness the efficacy of SSM, we introduce Efficient Vision Mamba (EfficientViM), a novel architecture built on hidden state mixer-based state space duality (HSM-SSD) that efficiently captures global dependencies with further reduced computational cost. With the observation that the runtime of the SSD layer is driven by the linear projections on the input sequences, we redesign the original SSD layer to perform the channel mixing operation within compressed hidden states in the HSM-SSD layer. Additionally, we propose multi-stage hidden state fusion to reinforce the representation power of hidden states and provide the design to alleviate the bottleneck caused by the memory-bound operations. As a result, the EfficientViM family achieves a new state-of-the-art speed-accuracy trade-off on ImageNet-1k, offering up to a 0.7% performance improvement over the second-best model SHViT with faster speed. Further, we observe significant improvements in throughput and accuracy compared to prior works, when scaling images or employing distillation training. Code is available at https://github.com/mlvlab/EfficientViM.

Jongha Kim, Jihwan Park, Jinyoung Park et al.

Visual Relationship Detection (VRD) has seen significant advancements with Transformer-based architectures recently. However, we identify two key limitations in a conventional label assignment for training Transformer-based VRD models, which is a process of mapping a ground-truth (GT) to a prediction. Under the conventional assignment, an unspecialized query is trained since a query is expected to detect every relation, which makes it difficult for a query to specialize in specific relations. Furthermore, a query is also insufficiently trained since a GT is assigned only to a single prediction, therefore near-correct or even correct predictions are suppressed by being assigned no relation as a GT. To address these issues, we propose Groupwise Query Specialization and Quality-Aware Multi-Assignment (SpeaQ). Groupwise Query Specialization trains a specialized query by dividing queries and relations into disjoint groups and directing a query in a specific query group solely toward relations in the corresponding relation group. Quality-Aware Multi-Assignment further facilitates the training by assigning a GT to multiple predictions that are significantly close to a GT in terms of a subject, an object, and the relation in between. Experimental results and analyses show that SpeaQ effectively trains specialized queries, which better utilize the capacity of a model, resulting in consistent performance gains with zero additional inference cost across multiple VRD models and benchmarks. Code is available at https://github.com/mlvlab/SpeaQ.

Juyeon Ko, Inho Kong, Dogyun Park et al.

Semantic image synthesis (SIS) is a task to generate realistic images corresponding to semantic maps (labels). However, in real-world applications, SIS often encounters noisy user inputs. To address this, we propose Stochastic Conditional Diffusion Model (SCDM), which is a robust conditional diffusion model that features novel forward and generation processes tailored for SIS with noisy labels. It enhances robustness by stochastically perturbing the semantic label maps through Label Diffusion, which diffuses the labels with discrete diffusion. Through the diffusion of labels, the noisy and clean semantic maps become similar as the timestep increases, eventually becoming identical at $t=T$. This facilitates the generation of an image close to a clean image, enabling robust generation. Furthermore, we propose a class-wise noise schedule to differentially diffuse the labels depending on the class. We demonstrate that the proposed method generates high-quality samples through extensive experiments and analyses on benchmark datasets, including a novel experimental setup simulating human errors during real-world applications. Code is available at https://github.com/mlvlab/SCDM.

Jaewon Chu, Jinyoung Park, Seunghun Lee et al.

Latent Bayesian optimization (LBO) approaches have successfully adopted Bayesian optimization over a continuous latent space by employing an encoder-decoder architecture to address the challenge of optimization in a high dimensional or discrete input space. LBO learns a surrogate model to approximate the black-box objective function in the latent space. However, we observed that most LBO methods suffer from the `misalignment problem`, which is induced by the reconstruction error of the encoder-decoder architecture. It hinders learning an accurate surrogate model and generating high-quality solutions. In addition, several trust region-based LBO methods select the anchor, the center of the trust region, based solely on the objective function value without considering the trust region`s potential to enhance the optimization process. To address these issues, we propose Inversion-based Latent Bayesian Optimization (InvBO), a plug-and-play module for LBO. InvBO consists of two components: an inversion method and a potential-aware trust region anchor selection. The inversion method searches the latent code that completely reconstructs the given target data. The potential-aware trust region anchor selection considers the potential capability of the trust region for better local optimization. Experimental results demonstrate the effectiveness of InvBO on nine real-world benchmarks, such as molecule design and arithmetic expression fitting tasks. Code is available at https://github.com/mlvlab/InvBO.

Dogyun Park, Sojin Lee, Sihyeon Kim et al.

Rectified flow and reflow procedures have significantly advanced fast generation by progressively straightening ordinary differential equation (ODE) flows. They operate under the assumption that image and noise pairs, known as couplings, can be approximated by straight trajectories with constant velocity. However, we observe that modeling with constant velocity and using reflow procedures have limitations in accurately learning straight trajectories between pairs, resulting in suboptimal performance in few-step generation. To address these limitations, we introduce Constant Acceleration Flow (CAF), a novel framework based on a simple constant acceleration equation. CAF introduces acceleration as an additional learnable variable, allowing for more expressive and accurate estimation of the ODE flow. Moreover, we propose two techniques to further improve estimation accuracy: initial velocity conditioning for the acceleration model and a reflow process for the initial velocity. Our comprehensive studies on toy datasets, CIFAR-10, and ImageNet 64x64 demonstrate that CAF outperforms state-of-the-art baselines for one-step generation. We also show that CAF dramatically improves few-step coupling preservation and inversion over Rectified flow. Code is available at \href{https://github.com/mlvlab/CAF}{https://github.com/mlvlab/CAF}.

Ji Soo Lee, Jongha Kim, Jeehye Na et al.

Despite the advancements of Video Large Language Models (VideoLLMs) in various tasks, they struggle with fine-grained temporal understanding, such as Dense Video Captioning (DVC). DVC is a complicated task of describing all events within a video while also temporally localizing them, which integrates multiple fine-grained tasks, including video segmentation, video captioning, and temporal video grounding. Previous VideoLLMs attempt to solve DVC in a single step, failing to utilize their reasoning capability. Moreover, previous training objectives for VideoLLMs do not fully reflect the evaluation metrics, therefore not providing supervision directly aligned to target tasks. To address such a problem, we propose a novel framework named VidChain comprised of Chain-of-Tasks (CoTasks) and Metric-based Direct Preference Optimization (M-DPO). CoTasks decompose a complex task into a sequence of sub-tasks, allowing VideoLLMs to leverage their reasoning capabilities more effectively. M-DPO aligns a VideoLLM with evaluation metrics, providing fine-grained supervision to each task that is well-aligned with metrics. Applied to two different VideoLLMs, VidChain consistently improves their fine-grained video understanding, thereby outperforming previous VideoLLMs on two different DVC benchmarks and also on the temporal video grounding task. Code is available at \url{https://github.com/mlvlab/VidChain}.

Ji Soo Lee, Byungoh Ko, Jaewon Cho et al.

In text-video retrieval, auxiliary captions are often used to enhance video understanding, bridging the gap between the modalities. While recent advances in multi-modal large language models (MLLMs) have enabled strong zero-shot caption generation, we observe that such captions tend to be generic and indistinguishable across visually similar videos, limiting their utility for fine-grained retrieval. Moreover, conventional captioning approaches are typically evaluated using language generation metrics, such as BLEU, which are not typically tailored for retrieval tasks that require making discriminative distinctions between candidates. To address this, we propose $\textbf{CaRe-DPO}$, a retrieval framework that directly optimizes caption generation using retrieval relevance scores. At its core is Dual-Group Direct Preference Optimization (DG-DPO), a novel learning strategy that supervises captioning by modeling preferences across groups of distinct video and caption pairs. In addition, we present an MLLM-based retrieval model that incorporates role-embeddings to better distinguish between textual inputs with different functional roles, such as an auxiliary caption and a text query. Through extensive experiments, we demonstrate that CaRe-DPO significantly enhances retrieval performance by effectively leveraging auxiliary knowledge to generate fine-grained captions for retrieval. Code is available at https://github.com/mlvlab/CaReDPO.

Dohwan Ko, Ji Soo Lee, Minhyuk Choi et al.

Text-Video Retrieval aims to find the most relevant text (or video) candidate given a video (or text) query from large-scale online databases. Recent work leverages multi-modal large language models (MLLMs) to improve retrieval, especially for long or complex query-candidate pairs. However, we observe that the naive application of MLLMs, i.e., retrieval based on candidate likelihood, introduces candidate prior bias, favoring candidates with inherently higher priors over those more relevant to the query. To this end, we propose a novel retrieval framework, Bidirectional Likelihood Estimation with MLLM (BLiM), which leverages both query and candidate likelihoods by training the model to generate text from a given video as well as video features from a given text. Furthermore, we introduce Candidate Prior Normalization (CPN), a simple yet effective training-free score calibration module designed to mitigate candidate prior bias in candidate likelihood. On four Text-Video Retrieval benchmarks, our BLiM equipped with CPN outperforms previous state-of-the-art models by 6.4 R@1 on average, effectively alleviating candidate prior bias and emphasizing query-candidate relevance. Our in-depth analysis across various multi-modal tasks beyond retrieval highlights the broad applicability of CPN which enhances visual understanding by reducing reliance on textual priors. Code is available at https://github.com/mlvlab/BLiM.

Sehyung Kim, Chanhyeong Yang, Jihwan Park et al.

Attribute classification is crucial for identifying specific characteristics within image regions. Vision-Language Models (VLMs) have been effective in zero-shot tasks by leveraging their general knowledge from large-scale datasets. Recent studies demonstrate that transformer-based models with class-wise queries can effectively address zero-shot multi-label classification. However, poor utilization of the relationship between seen and unseen attributes makes the model lack generalizability. Additionally, attribute classification generally involves many attributes, making maintaining the model's scalability difficult. To address these issues, we propose Super-class guided transFormer (SugaFormer), a novel framework that leverages super-classes to enhance scalability and generalizability for zero-shot attribute classification. SugaFormer employs Super-class Query Initialization (SQI) to reduce the number of queries, utilizing common semantic information from super-classes, and incorporates Multi-context Decoding (MD) to handle diverse visual cues. To strengthen generalizability, we introduce two knowledge transfer strategies that utilize VLMs. During training, Super-class guided Consistency Regularization (SCR) aligns model's features with VLMs using super-class guided prompts, and during inference, Zero-shot Retrieval-based Score Enhancement (ZRSE) refines predictions for unseen attributes. Extensive experiments demonstrate that SugaFormer achieves state-of-the-art performance across three widely-used attribute classification benchmarks under zero-shot, and cross-dataset transfer settings. Our code is available at https://github.com/mlvlab/SugaFormer.

Bumsoo Kim, Taeho Choi, Jaewoo Kang et al.

Recent advances in deep neural networks have achieved significant progress in detecting individual objects from an image. However, object detection is not sufficient to fully understand a visual scene. Towards a deeper visual understanding, the interactions between objects, especially humans and objects are essential. Most prior works have obtained this information with a bottom-up approach, where the objects are first detected and the interactions are predicted sequentially by pairing the objects. This is a major bottleneck in HOI detection inference time. To tackle this problem, we propose UnionDet, a one-stage meta-architecture for HOI detection powered by a novel union-level detector that eliminates this additional inference stage by directly capturing the region of interaction. Our one-stage detector for human-object interaction shows a significant reduction in interaction prediction time 4x~14x while outperforming state-of-the-art methods on two public datasets: V-COCO and HICO-DET.

Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao et al.

Multimodal Large Language Models (MLLMs) have shown promising progress in understanding and analyzing video content. However, processing long videos remains a significant challenge constrained by LLM's context size. To address this limitation, we propose LongVU, a spatiotemporal adaptive compression mechanism thats reduces the number of video tokens while preserving visual details of long videos. Our idea is based on leveraging cross-modal query and inter-frame dependencies to adaptively reduce temporal and spatial redundancy in videos. Specifically, we leverage DINOv2 features to remove redundant frames that exhibit high similarity. Then we utilize text-guided cross-modal query for selective frame feature reduction. Further, we perform spatial token reduction across frames based on their temporal dependencies. Our adaptive compression strategy effectively processes a large number of frames with little visual information loss within given context length. Our LongVU consistently surpass existing methods across a variety of video understanding benchmarks, especially on hour-long video understanding tasks such as VideoMME and MLVU. Given a light-weight LLM, our LongVU also scales effectively into a smaller size with state-of-the-art video understanding performance.

Jongha Kim, Minseong Bae, Sanghyeok Lee et al.

Table images present unique challenges for effective and efficient understanding due to the need for question-specific focus and the presence of redundant background regions. Existing Multimodal Large Language Model (MLLM) approaches often overlook these characteristics, resulting in uninformative and redundant visual representations. To address these issues, we aim to generate visual features that are both informative and compact to improve table understanding. We first propose progressive question conditioning, which injects the question into Vision Transformer layers with gradually increasing frequency, considering each layer's capacity to handle additional information, to generate question-aware visual features. To reduce redundancy, we introduce a pruning strategy that discards background tokens, thereby improving efficiency. To mitigate information loss from pruning, we further propose token focusing, a training strategy that encourages the model to concentrate essential information in the retained tokens. By combining these approaches, we present TabFlash, an efficient and effective MLLM for table understanding. TabFlash achieves state-of-the-art performance, outperforming both open-source and proprietary MLLMs, while requiring 27% less FLOPs and 30% less memory usage compared to the second-best MLLM.

Jaewon Chu, Seunghun Lee, Hyunwoo J. Kim

Large language models (LLMs) have achieved remarkable success across diverse domains, due to their strong instruction-following capabilities. This has led to increasing interest in optimizing instructions for black-box LLMs, whose internal parameters are inaccessible but widely used due to their strong performance. To optimize instructions for black-box LLMs, recent methods employ white-box LLMs to generate candidate instructions from optimized soft prompts. However, white-box LLMs often map different soft prompts to the same instruction, leading to redundant queries. While previous studies regarded this many-to-one mapping as a structure that hinders optimization efficiency, we reinterpret it as a useful prior knowledge that can accelerate the optimization. To this end, we introduce PREimage-informed inSTruction Optimization (PRESTO), a novel framework that leverages the preimage structure of soft prompts for efficient optimization. PRESTO consists of three key components: (1) score sharing, which shares the evaluation score with all soft prompts in a preimage; (2) preimage-based initialization, which selects initial data points that maximize search space coverage using preimage information; and (3) score consistency regularization, which enforces prediction consistency within each preimage. By leveraging preimages, PRESTO achieves the effect of effectively obtaining 14 times more scored data under the same query budget, resulting in more efficient optimization. Experimental results on 33 instruction optimization tasks demonstrate the superior performance of PRESTO. Code is available at https://github.com/mlvlab/PRESTO