Suk-Ju Kang

Hyunwoo Yu, Yubin Cho, Beoungwoo Kang et al.

We present an Encoder-Decoder Attention Transformer, EDAFormer, which consists of the Embedding-Free Transformer (EFT) encoder and the all-attention decoder leveraging our Embedding-Free Attention (EFA) structure. The proposed EFA is a novel global context modeling mechanism that focuses on functioning the global non-linearity, not the specific roles of the query, key and value. For the decoder, we explore the optimized structure for considering the globality, which can improve the semantic segmentation performance. In addition, we propose a novel Inference Spatial Reduction (ISR) method for the computational efficiency. Different from the previous spatial reduction attention methods, our ISR method further reduces the key-value resolution at the inference phase, which can mitigate the computation-performance trade-off gap for the efficient semantic segmentation. Our EDAFormer shows the state-of-the-art performance with the efficient computation compared to the existing transformer-based semantic segmentation models in three public benchmarks, including ADE20K, Cityscapes and COCO-Stuff. Furthermore, our ISR method reduces the computational cost by up to 61% with minimal mIoU performance degradation on Cityscapes dataset. The code is available at https://github.com/hyunwoo137/EDAFormer.

Beoungwoo Kang, Seunghun Moon, Yubin Cho et al.

Beyond the Transformer, it is important to explore how to exploit the capacity of the MetaFormer, an architecture that is fundamental to the performance improvements of the Transformer. Previous studies have exploited it only for the backbone network. Unlike previous studies, we explore the capacity of the Metaformer architecture more extensively in the semantic segmentation task. We propose a powerful semantic segmentation network, MetaSeg, which leverages the Metaformer architecture from the backbone to the decoder. Our MetaSeg shows that the MetaFormer architecture plays a significant role in capturing the useful contexts for the decoder as well as for the backbone. In addition, recent segmentation methods have shown that using a CNN-based backbone for extracting the spatial information and a decoder for extracting the global information is more effective than using a transformer-based backbone with a CNN-based decoder. This motivates us to adopt the CNN-based backbone using the MetaFormer block and design our MetaFormer-based decoder, which consists of a novel self-attention module to capture the global contexts. To consider both the global contexts extraction and the computational efficiency of the self-attention for semantic segmentation, we propose a Channel Reduction Attention (CRA) module that reduces the channel dimension of the query and key into the one dimension. In this way, our proposed MetaSeg outperforms the previous state-of-the-art methods with more efficient computational costs on popular semantic segmentation and a medical image segmentation benchmark, including ADE20K, Cityscapes, COCO-stuff, and Synapse. The code is available at https://github.com/hyunwoo137/MetaSeg.

Junho Park, Kyeongbo Kong, Suk-Ju Kang

Recently, there has been a significant amount of research conducted on 3D hand reconstruction to use various forms of human-computer interaction. However, 3D hand reconstruction in the wild is challenging due to extreme lack of in-the-wild 3D hand datasets. Especially, when hands are in complex pose such as interacting hands, the problems like appearance similarity, self-handed occclusion and depth ambiguity make it more difficult. To overcome these issues, we propose AttentionHand, a novel method for text-driven controllable hand image generation. Since AttentionHand can generate various and numerous in-the-wild hand images well-aligned with 3D hand label, we can acquire a new 3D hand dataset, and can relieve the domain gap between indoor and outdoor scenes. Our method needs easy-to-use four modalities (i.e, an RGB image, a hand mesh image from 3D label, a bounding box, and a text prompt). These modalities are embedded into the latent space by the encoding phase. Then, through the text attention stage, hand-related tokens from the given text prompt are attended to highlight hand-related regions of the latent embedding. After the highlighted embedding is fed to the visual attention stage, hand-related regions in the embedding are attended by conditioning global and local hand mesh images with the diffusion-based pipeline. In the decoding phase, the final feature is decoded to new hand images, which are well-aligned with the given hand mesh image and text prompt. As a result, AttentionHand achieved state-of-the-art among text-to-hand image generation models, and the performance of 3D hand mesh reconstruction was improved by additionally training with hand images generated by AttentionHand.

Yubin Cho, Hyunwoo Yu, Suk-ju Kang

Referring segmentation aims to segment a target object related to a natural language expression. Key challenges of this task are understanding the meaning of complex and ambiguous language expressions and determining the relevant regions in the image with multiple objects by referring to the expression. Recent models have focused on the early fusion with the language features at the intermediate stage of the vision encoder, but these approaches have a limitation that the language features cannot refer to the visual information. To address this issue, this paper proposes a novel architecture, Cross-aware early fusion with stage-divided Vision and Language Transformer encoders (CrossVLT), which allows both language and vision encoders to perform the early fusion for improving the ability of the cross-modal context modeling. Unlike previous methods, our method enables the vision and language features to refer to each other's information at each stage to mutually enhance the robustness of both encoders. Furthermore, unlike the conventional scheme that relies solely on the high-level features for the cross-modal alignment, we introduce a feature-based alignment scheme that enables the low-level to high-level features of the vision and language encoders to engage in the cross-modal alignment. By aligning the intermediate cross-modal features in all encoder stages, this scheme leads to effective cross-modal fusion. In this way, the proposed approach is simple but effective for referring image segmentation, and it outperforms the previous state-of-the-art methods on three public benchmarks.

Ginam Kim, Hyunsung Kim, Joseph Kihoon Kim et al.

Recently, along with interest in autonomous vehicles, the importance of monitoring systems for both drivers and passengers inside vehicles has been increasing. This paper proposes a novel in-vehicle monitoring system the combines 3D pose estimation, seat-belt segmentation, and seat-belt status classification networks. Our system outputs various information necessary for monitoring by accurately considering the data characteristics of the in-vehicle environment. Specifically, the proposed 3D pose estimation directly estimates the absolute coordinates of keypoints for a driver and passengers, and the proposed seat-belt segmentation is implemented by applying a structure based on the feature pyramid. In addition, we propose a classification task to distinguish between normal and abnormal states of wearing a seat belt using results that combine 3D pose estimation with seat-belt segmentation. These tasks can be learned simultaneously and operate in real-time. Our method was evaluated on a private dataset we newly created and annotated. The experimental results show that our method has significantly high performance that can be applied directly to real in-vehicle monitoring systems.

Yubin Cho, Hyunwoo Yu, Kyeongbo Kong et al.

Referring Image Segmentation (RIS) aims to segment a target object described by a natural language expression. Existing methods have evolved by leveraging the vision information into the language tokens. To more effectively exploit visual contexts for fine-grained segmentation, we propose a novel Visual Informative Part Attention (VIPA) framework for referring image segmentation. VIPA leverages the informative parts of visual contexts, called a visual expression, which can effectively provide the structural and semantic visual target information to the network. This design reduces high-variance cross-modal projection and enhances semantic consistency in an attention mechanism of the referring image segmentation. We also design a visual expression generator (VEG) module, which retrieves informative visual tokens via local-global linguistic context cues and refines the retrieved tokens for reducing noise information and sharing informative visual attributes. This module allows the visual expression to consider comprehensive contexts and capture semantic visual contexts of informative regions. In this way, our framework enables the network's attention to robustly align with the fine-grained regions of interest. Extensive experiments and visual analysis demonstrate the effectiveness of our approach. Our VIPA outperforms the existing state-of-the-art methods on four public RIS benchmarks.

Gyeongrok Oh, Youngdong Jang, Jonghyun Choi et al.

Dominant paradigms for 4D LiDAR panoptic segmentation are usually required to train deep neural networks with large superimposed point clouds or design dedicated modules for instance association. However, these approaches perform redundant point processing and consequently become computationally expensive, yet still overlook the rich geometric priors inherently provided by raw point clouds. To this end, we introduce ICP-4D, a simple yet effective training-free framework that unifies spatial and temporal reasoning through geometric relations among instance-level point sets. Specifically, we apply the Iterative Closest Point (ICP) algorithm to directly associate temporally consistent instances by aligning the source and target point sets through the estimated transformation. To stabilize association under noisy instance predictions, we introduce a Sinkhorn-based soft matching. This exploits the underlying instance distribution to obtain accurate point-wise correspondences, resulting in robust geometric alignment. Furthermore, our carefully designed pipeline, which considers three instance types-static, dynamic, and missing-offers computational efficiency and occlusion-aware matching. Our extensive experiments across both SemanticKITTI and panoptic nuScenes demonstrate that our method consistently outperforms state-of-the-art approaches, even without additional training or extra point cloud inputs.

Kyeongbo Kong, Kyunghun Kim, Suk-Ju Kang

Conditional generative adversarial networks (cGANs) have demonstrated remarkable success due to their class-wise controllability and superior quality for complex generation tasks. Typical cGANs solve the joint distribution matching problem by decomposing two easier sub-problems: marginal matching and conditional matching. In this paper, we proposes a simple but effective training methodology, selective focusing learning, which enforces the discriminator and generator to learn easy samples of each class rapidly while maintaining diversity. Our key idea is to selectively apply conditional and joint matching for the data in each mini-batch.Specifically, we first select the samples with the highest scores when sorted using the conditional term of the discriminator outputs (real and generated samples). Then we optimize the model using the selected samples with only conditional matching and the other samples with joint matching. From our toy experiments, we found that it is the best to apply only conditional matching to certain samples due to the content-aware optimization of the discriminator. We conducted experiments on ImageNet (64x64 and 128x128), CIFAR-10, CIFAR-100 datasets, and Mixture of Gaussian, noisy label settings to demonstrate that the proposed method can substantially (up to 35.18% in terms of FID) improve all indicators with 10 independent trials. Code is available at https://github.com/pnu-cvsp/Enhancing-Stability-in-Training-Conditional-GAN-via-Selective-Data-Matching.

Taewoo Kim, Suwoong Yeom, Jaehyun Pyun et al.

Reconstructing dynamic scenes with complex human-object interactions is a fundamental challenge in computer vision and graphics. Existing Gaussian Splatting methods either rely on human pose priors while neglecting dynamic objects, or approximate all motions within a single field, limiting their ability to capture interaction-rich dynamics. To address this gap, we propose Human-Object Interaction Gaussian Splatting (HOIGS), which explicitly models interaction-induced deformation between humans and objects through a cross-attention-based HOI module. Distinct deformation baselines are employed to extract features: HexPlane for humans and Cubic Hermite Spline (CHS) for objects. By integrating these heterogeneous features, HOIGS effectively captures interdependent motions and improves deformation estimation in scenarios involving occlusion, contact, and object manipulation. Comprehensive experiments on multiple datasets demonstrate that our method consistently outperforms state-of-the-art human-centric and 4D Gaussian approaches, highlighting the importance of explicitly modeling human-object interactions for high-fidelity reconstruction.

Jihyun Kim, Junho Park, Kyeongbo Kong et al.

We present Programmable-Room, a framework which interactively generates and edits a 3D room mesh, given natural language instructions. For precise control of a room's each attribute, we decompose the challenging task into simpler steps such as creating plausible 3D coordinates for room meshes, generating panorama images for the texture, constructing 3D meshes by integrating the coordinates and panorama texture images, and arranging furniture. To support the various decomposed tasks with a unified framework, we incorporate visual programming (VP). VP is a method that utilizes a large language model (LLM) to write a Python-like program which is an ordered list of necessary modules for the various tasks given in natural language. We develop most of the modules. Especially, for the texture generating module, we utilize a pretrained large-scale diffusion model to generate panorama images conditioned on text and visual prompts (i.e., layout, depth, and semantic map) simultaneously. Specifically, we enhance the panorama image generation quality by optimizing the training objective with a 1D representation of a panorama scene obtained from bidirectional LSTM. We demonstrate Programmable-Room's flexibility in generating and editing 3D room meshes, and prove our framework's superiority to an existing model quantitatively and qualitatively. Project page is available in https://jihyun0510.github.io/Programmable_Room_Page/.

Jouwon Song, Kyeongbo Kong, Ye-In Park et al.

Anomaly segmentation, which localizes defective areas, is an important component in large-scale industrial manufacturing. However, most recent researches have focused on anomaly detection. This paper proposes a novel anomaly segmentation network (AnoSeg) that can directly generate an accurate anomaly map using self-supervised learning. For highly accurate anomaly segmentation, the proposed AnoSeg considers three novel techniques: Anomaly data generation based on hard augmentation, self-supervised learning with pixel-wise and adversarial losses, and coordinate channel concatenation. First, to generate synthetic anomaly images and reference masks for normal data, the proposed method uses hard augmentation to change the normal sample distribution. Then, the proposed AnoSeg is trained in a self-supervised learning manner from the synthetic anomaly data and normal data. Finally, the coordinate channel, which represents the pixel location information, is concatenated to an input of AnoSeg to consider the positional relationship of each pixel in the image. The estimated anomaly map can also be utilized to improve the performance of anomaly detection. Our experiments show that the proposed method outperforms the state-of-the-art anomaly detection and anomaly segmentation methods for the MVTec AD dataset. In addition, we compared the proposed method with the existing methods through the intersection over union (IoU) metric commonly used in segmentation tasks and demonstrated the superiority of our method for anomaly segmentation.

Jae-hun Shim, Kyeongbo Kong, Suk-Ju Kang

Neural architecture search (NAS), an important branch of automatic machine learning, has become an effective approach to automate the design of deep learning models. However, the major issue in NAS is how to reduce the large search time imposed by the heavy computational burden. While most recent approaches focus on pruning redundant sets or developing new search methodologies, this paper attempts to formulate the problem based on the data curation manner. Our key strategy is to search the architecture using summarized data distribution, i.e., core-set. Typically, many NAS algorithms separate searching and training stages, and the proposed core-set methodology is only used in search stage, thus their performance degradation can be minimized. In our experiments, we were able to save overall computational time from 30.8 hours to 3.5 hours, 8.8x reduction, on a single RTX 3090 GPU without sacrificing accuracy.

Jou Won Song, Kyeongbo Kong, Ye In Park et al.

Anomaly detection is a task that recognizes whether an input sample is included in the distribution of a target normal class or an anomaly class. Conventional generative adversarial network (GAN)-based methods utilize an entire image including foreground and background as an input. However, in these methods, a useless region unrelated to the normal class (e.g., unrelated background) is learned as normal class distribution, thereby leading to false detection. To alleviate this problem, this paper proposes a novel two-stage network consisting of an attention network and an anomaly detection GAN (ADGAN). The attention network generates an attention map that can indicate the region representing the normal class distribution. To generate an accurate attention map, we propose the attention loss and the adversarial anomaly loss based on synthetic anomaly samples generated from hard augmentation. By applying the attention map to an image feature map, ADGAN learns the normal class distribution from which the useless region is removed, and it is possible to greatly reduce the problem difficulty of the anomaly detection task. Additionally, the estimated attention map can be used for anomaly segmentation because it can distinguish between normal and anomaly regions. As a result, the proposed method outperforms the state-of-the-art anomaly detection and anomaly segmentation methods for widely used datasets.

Kyunghun Kim, Yeohun Yun, Keon-Woo Kang et al.

Image outpainting is a very intriguing problem as the outside of a given image can be continuously filled by considering as the context of the image. This task has two main challenges. The first is to maintain the spatial consistency in contents of generated regions and the original input. The second is to generate a high-quality large image with a small amount of adjacent information. Conventional image outpainting methods generate inconsistent, blurry, and repeated pixels. To alleviate the difficulty of an outpainting problem, we propose a novel image outpainting method using bidirectional boundary region rearrangement. We rearrange the image to benefit from the image inpainting task by reflecting more directional information. The bidirectional boundary region rearrangement enables the generation of the missing region using bidirectional information similar to that of the image inpainting task, thereby generating the higher quality than the conventional methods using unidirectional information. Moreover, we use the edge map generator that considers images as original input with structural information and hallucinates the edges of unknown regions to generate the image. Our proposed method is compared with other state-of-the-art outpainting and inpainting methods both qualitatively and quantitatively. We further compared and evaluated them using BRISQUE, one of the No-Reference image quality assessment (IQA) metrics, to evaluate the naturalness of the output. The experimental results demonstrate that our method outperforms other methods and generates new images with 360°panoramic characteristics.

Jung Hee Kim, Siyeong Lee, Suk-Ju Kang

Recently, high dynamic range (HDR) image reconstruction based on the multiple exposure stack from a given single exposure utilizes a deep learning framework to generate high-quality HDR images. These conventional networks focus on the exposure transfer task to reconstruct the multi-exposure stack. Therefore, they often fail to fuse the multi-exposure stack into a perceptually pleasant HDR image as the inversion artifacts occur. We tackle the problem in stack reconstruction-based methods by proposing a novel framework with a fully differentiable high dynamic range imaging (HDRI) process. By explicitly using the loss, which compares the network's output with the ground truth HDR image, our framework enables a neural network that generates the multiple exposure stack for HDRI to train stably. In other words, our differentiable HDR synthesis layer helps the deep neural network to train to create multi-exposure stacks while reflecting the precise correlations between multi-exposure images in the HDRI process. In addition, our network uses the image decomposition and the recursive process to facilitate the exposure transfer task and to adaptively respond to recursion frequency. The experimental results show that the proposed network outperforms the state-of-the-art quantitative and qualitative results in terms of both the exposure transfer tasks and the whole HDRI process.

Saehyun Ahn, Jung-Woo Chang, Suk-Ju Kang

Deformable convolutional networks have demonstrated outstanding performance in object recognition tasks with an effective feature extraction. Unlike standard convolution, the deformable convolution decides the receptive field size using dynamically generated offsets, which leads to an irregular memory access. Especially, the memory access pattern varies both spatially and temporally, making static optimization ineffective. Thus, a naive implementation would lead to an excessive memory footprint. In this paper, we present a novel approach to accelerate deformable convolution on FPGA. First, we propose a novel training method to reduce the size of the receptive field in the deformable convolutional layer without compromising accuracy. By optimizing the receptive field, we can compress the maximum size of the receptive field by 12.6 times. Second, we propose an efficient systolic architecture to maximize its efficiency. We then implement our design on FPGA to support the optimized dataflow. Experimental results show that our accelerator achieves up to 17.25 times speedup over the state-of-the-art accelerator.

Jung-Woo Chang, Saehyun Ahn, Keon-Woo Kang et al.

Generative adversarial networks (GANs) have shown excellent performance in image and speech applications. GANs create impressive data primarily through a new type of operator called deconvolution (DeConv) or transposed convolution (Conv). To implement the DeConv layer in hardware, the state-of-the-art accelerator reduces the high computational complexity via the DeConv-to-Conv conversion and achieves the same results. However, there is a problem that the number of filters increases due to this conversion. Recently, Winograd minimal filtering has been recognized as an effective solution to improve the arithmetic complexity and resource efficiency of the Conv layer. In this paper, we propose an efficient Winograd DeConv accelerator that combines these two orthogonal approaches on FPGAs. Firstly, we introduce a new class of fast algorithm for DeConv layers using Winograd minimal filtering. Since there are regular sparse patterns in Winograd filters, we further amortize the computational complexity by skipping zero weights. Secondly, we propose a new dataflow to prevent resource underutilization by reorganizing the filter layout in the Winograd domain. Finally, we propose an efficient architecture for implementing Winograd DeConv by designing the line buffer and exploring the design space. Experimental results on various GANs show that our accelerator achieves up to 1.78x~8.38x speedup over the state-of-the-art DeConv accelerators.

Siyeong Lee, Gwon Hwan An, Suk-Ju Kang

In this paper, we propose a novel deep neural network model that reconstructs a high dynamic range (HDR) image from a single low dynamic range (LDR) image. The proposed model is based on a convolutional neural network composed of dilated convolutional layers, and infers LDR images with various exposures and illumination from a single LDR image of the same scene. Then, the final HDR image can be formed by merging these inference results. It is relatively easy for the proposed method to find the mapping between the LDR and an HDR with a different bit depth because of the chaining structure inferring the relationship between the LDR images with brighter (or darker) exposures from a given LDR image. The method not only extends the range, but also has the advantage of restoring the light information of the actual physical world. For the HDR images obtained by the proposed method, the HDR-VDP2 Q score, which is the most popular evaluation metric for HDR images, was 56.36 for a display with a 1920$\times$1200 resolution, which is an improvement of 6 compared with the scores of conventional algorithms. In addition, when comparing the peak signal-to-noise ratio values for tone mapped HDR images generated by the proposed and conventional algorithms, the average value obtained by the proposed algorithm is 30.86 dB, which is 10 dB higher than those obtained by the conventional algorithms.