Chih–Chung Hsu

Benjamin Kiefer, Matej Kristan, Janez Perš et al.

The 1$^{\text{st}}$ Workshop on Maritime Computer Vision (MaCVi) 2023 focused on maritime computer vision for Unmanned Aerial Vehicles (UAV) and Unmanned Surface Vehicle (USV), and organized several subchallenges in this domain: (i) UAV-based Maritime Object Detection, (ii) UAV-based Maritime Object Tracking, (iii) USV-based Maritime Obstacle Segmentation and (iv) USV-based Maritime Obstacle Detection. The subchallenges were based on the SeaDronesSee and MODS benchmarks. This report summarizes the main findings of the individual subchallenges and introduces a new benchmark, called SeaDronesSee Object Detection v2, which extends the previous benchmark by including more classes and footage. We provide statistical and qualitative analyses, and assess trends in the best-performing methodologies of over 130 submissions. The methods are summarized in the appendix. The datasets, evaluation code and the leaderboard are publicly available at https://seadronessee.cs.uni-tuebingen.de/macvi.

Chih-Chung Hsu, Xin-Di Ma, Wo-Ting Liao et al.

Existing attention accelerators often trade exact softmax semantics, depend on fused Tensor Core kernels, or incur sequential depth that limits FP32 throughput on long sequences. We present \textbf{ELSA}, an algorithmic reformulation of online softmax attention that (i)~preserves exact softmax semantics in real arithmetic with a \emph{provable} $\mathcal{O}(u\log n)$ FP32 relative error bound; (ii)~casts the online softmax update as a prefix scan over an associative monoid $(m,S,W)$, yielding $O(n)$ extra memory and $O(\log n)$ parallel depth; and (iii)~is Tensor-Core independent, implemented in Triton and CUDA C++, and deployable as a \emph{drop-in replacement} requiring no retraining or weight modification. Unlike FlashAttention-2/3, which rely on HMMA/GMMA Tensor Core instructions and provide no compatible FP32 path, ELSA operates identically on A100s and resource-constrained edge devices such as Jetson TX2 -- making it the only hardware-agnostic exact-attention kernel that reduces parallel depth to $O(\log n)$ at full precision. On A100 FP32 benchmarks (1K--16K tokens), ELSA delivers $1.3$--$3.5\times$ speedup over memory-efficient SDPA and $1.97$--$2.27\times$ on BERT; on Jetson TX2, ELSA achieves $1.5$--$1.6\times$ over Math (64--900 tokens), with $17.8$--$20.2\%$ throughput gains under LLaMA-13B offloading at $\ge$32K. In FP16, ELSA approaches hardware-fused baselines at long sequences while retaining full FP32 capability, offering a unified kernel for high-precision inference across platforms. Our code and implementation are available at https://github.com/ming053l/ELSA.

Yu-Fan Lin, Ching-Heng Cheng, Bo-Cheng Qiu et al.

Fusarium Head Blight (FHB) is a serious fungal disease affecting wheat (including durum), barley, oats, other small cereal grains, and corn. Effective monitoring and accurate detection of FHB are crucial to ensuring stable and reliable food security. Traditionally, trained agronomists and surveyors perform manual identification, a method that is labor-intensive, impractical, and challenging to scale. With the advancement of deep learning and Hyper-spectral Imaging (HSI) and Remote Sensing (RS) technologies, employing deep learning, particularly Convolutional Neural Networks (CNNs), has emerged as a promising solution. Notably, wheat infected with serious FHB may exhibit significant differences on the spectral compared to mild FHB one, which is particularly advantageous for hyperspectral image-based methods. In this study, we propose a self-unsupervised classification method based on HSI endmember extraction strategy and top-K bands selection, designed to analyze material signatures in HSIs to derive discriminative feature representations. This approach does not require expensive device or complicate algorithm design, making it more suitable for practical uses. Our method has been effectively validated in the Beyond Visible Spectrum: AI for Agriculture Challenge 2024. The source code is easy to reproduce and available at {https://github.com/VanLinLin/Automated-Crop-Disease-Diagnosis-from-Hyperspectral-Imagery-3rd}.

Chih-Chung Hsu, Chia-Ming Lee, Yi-Shiuan Chou

In recent years, Vision Transformer-based approaches for low-level vision tasks have achieved widespread success. Unlike CNN-based models, Transformers are more adept at capturing long-range dependencies, enabling the reconstruction of images utilizing non-local information. In the domain of super-resolution, Swin-transformer-based models have become mainstream due to their capability of global spatial information modeling and their shifting-window attention mechanism that facilitates the interchange of information between different windows. Many researchers have enhanced model performance by expanding the receptive fields or designing meticulous networks, yielding commendable results. However, we observed that it is a general phenomenon for the feature map intensity to be abruptly suppressed to small values towards the network's end. This implies an information bottleneck and a diminishment of spatial information, implicitly limiting the model's potential. To address this, we propose the Dense-residual-connected Transformer (DRCT), aimed at mitigating the loss of spatial information and stabilizing the information flow through dense-residual connections between layers, thereby unleashing the model's potential and saving the model away from information bottleneck. Experiment results indicate that our approach surpasses state-of-the-art methods on benchmark datasets and performs commendably at the NTIRE-2024 Image Super-Resolution (x4) Challenge. Our source code is available at https://github.com/ming053l/DRCT

Bo-Cheng Qiu, Fang-Ying Lin, Ming-Han Sun et al.

Capsule endoscopy event detection is challenging because clinically relevant findings are sparse, visually heterogeneous, and evaluated at the event level rather than by frame accuracy. We propose VISTA, a metric-aligned multi-backbone framework for the RAREVISION task. VISTA combines EndoFM-LV for temporal context and DINOv3 ViTL/16 for frame-level visual semantics, followed by a Diverse Head Ensemble (DHE), Validation-Guided Weighted Fusion (VGWF), and Anatomy-Aware Temporal Event Decoding (ATED). The original official submission achieved hidden-test temporal mAP@0.5 of 0.3530 and mAP@0.95 of 0.3235. After the competition, extending local threshold refinement with a global coarse search improved performance to 0.3726 mAP@0.5 and 0.3431 mAP@0.95, ranking Team ACVLab second in the post-competition evaluation.

Bo-Cheng Qiu, Yu-Fan Lin, Yu-Zhe Pien et al.

Capsule endoscopy event detection is challenging because diagnostically relevant findings are sparse, visually heterogeneous, and embedded in long, noisy video streams, while evaluation is performed at the event level rather than by frame accuracy alone. We therefore formulate the RARE-VISION task as a metric-aligned event detection problem instead of a purely frame-wise classification task. Our framework combines two complementary backbones, EndoFM-LV for local temporal context and DINOv3 ViT-L/16 for strong frame-level visual semantics, followed by a Diverse Head Ensemble, Validation-Guided Hierarchical Fusion, and Anatomy-Aware Temporal Event Decoding. The fusion stage uses validation-derived class-wise model weighting, backbone weighting, and probability calibration, while the decoding stage applies temporal smoothing, anatomical constraints, threshold refinement, and per-label event generation to produce stable event predictions. Validation ablations indicate that complementary backbones, validation-guided fusion, and anatomy-aware temporal decoding all contribute to event-level performance. On the official hidden test set, the proposed method achieved an overall temporal mAP@0.5 of 0.3530 and temporal mAP@0.95 of 0.3235.

Chih-Chung Hsu, Chi-Han Tsai, Guan-Lin Chen et al.

Computed tomography (CT) imaging could be very practical for diagnosing various diseases. However, the nature of the CT images is even more diverse since the resolution and number of the slices of a CT scan are determined by the machine and its settings. Conventional deep learning models are hard to tickle such diverse data since the essential requirement of the deep neural network is the consistent shape of the input data. In this paper, we propose a novel, effective, two-step-wise approach to tickle this issue for COVID-19 symptom classification thoroughly. First, the semantic feature embedding of each slice for a CT scan is extracted by conventional backbone networks. Then, we proposed a long short-term memory (LSTM) and Transformer-based sub-network to deal with temporal feature learning, leading to spatiotemporal feature representation learning. In this fashion, the proposed two-step LSTM model could prevent overfitting, as well as increase performance. Comprehensive experiments reveal that the proposed two-step method not only shows excellent performance but also could be compensated for each other. More specifically, the two-step LSTM model has a lower false-negative rate, while the 2-step Swin model has a lower false-positive rate. In summary, it is suggested that the model ensemble could be adopted for more stable and promising performance in real-world applications.

Chih-Chung Hsu, Chih-Yu Jian, Chia-Ming Lee et al.

This paper investigates the application of deep learning models for lung Computed Tomography (CT) image analysis. Traditional deep learning frameworks encounter compatibility issues due to variations in slice numbers and resolutions in CT images, which stem from the use of different machines. Commonly, individual slices are predicted and subsequently merged to obtain the final result; however, this approach lacks slice-wise feature learning and consequently results in decreased performance. We propose a novel slice selection method for each CT dataset to address this limitation, effectively filtering out uncertain slices and enhancing the model's performance. Furthermore, we introduce a spatial-slice feature learning (SSFL) technique\cite{hsu2022} that employs a conventional and efficient backbone model for slice feature training, followed by extracting one-dimensional data from the trained model for COVID and non-COVID classification using a dedicated classification model. Leveraging these experimental steps, we integrate one-dimensional features with multiple slices for channel merging and employ a 2D convolutional neural network (CNN) model for classification. In addition to the aforementioned methods, we explore various high-performance classification models, ultimately achieving promising results.

Chih-Chung Hsu, Chia-Ming Lee, Yang Fan Chiang et al.

Conventional Computed Tomography (CT) imaging recognition faces two significant challenges: (1) There is often considerable variability in the resolution and size of each CT scan, necessitating strict requirements for the input size and adaptability of models. (2) CT-scan contains large number of out-of-distribution (OOD) slices. The crucial features may only be present in specific spatial regions and slices of the entire CT scan. How can we effectively figure out where these are located? To deal with this, we introduce an enhanced Spatial-Slice Feature Learning (SSFL++) framework specifically designed for CT scan. It aim to filter out a OOD data within whole CT scan, enabling our to select crucial spatial-slice for analysis by reducing 70% redundancy totally. Meanwhile, we proposed Kernel-Density-based slice Sampling (KDS) method to improve the stability when training and inference stage, therefore speeding up the rate of convergence and boosting performance. As a result, the experiments demonstrate the promising performance of our model using a simple EfficientNet-2D (E2D) model, even with only 1% of the training data. The efficacy of our approach has been validated on the COVID-19-CT-DB datasets provided by the DEF-AI-MIA workshop, in conjunction with CVPR 2024. Our source code is available at https://github.com/ming053l/E2D

Chia-Ming Lee, Ching-Heng Cheng, Yu-Fan Lin et al.

Recent advances in All-in-One (AiO) RGB image restoration have demonstrated the effectiveness of prompt learning in handling multiple degradations within a single model. However, extending these approaches to hyperspectral image (HSI) restoration is challenging due to the domain gap between RGB and HSI features, information loss in visual prompts under severe composite degradations, and difficulties in capturing HSI-specific degradation patterns via text prompts. In this paper, we propose PromptHSI, the first universal AiO HSI restoration framework that addresses these challenges. By incorporating frequency-aware feature modulation, which utilizes frequency analysis to narrow down the restoration search space and employing vision-language model (VLM)-guided prompt learning, our approach decomposes text prompts into intensity and bias controllers that effectively guide the restoration process while mitigating domain discrepancies. Extensive experiments demonstrate that our unified architecture excels at both fine-grained recovery and global information restoration across diverse degradation scenarios, highlighting its significant potential for practical remote sensing applications. The source code is available at https://github.com/chingheng0808/PromptHSI.

Yu-Fan Lin, Chia-Ming Lee, Chih-Chung Hsu

Shadows are a common factor degrading image quality. Single-image shadow removal (SR), particularly under challenging indirect illumination, is hampered by non-uniform content degradation and inherent ambiguity. Consequently, traditional methods often fail to simultaneously recover intra-shadow details and maintain sharp boundaries, resulting in inconsistent restoration and blurring that negatively affect both downstream applications and the overall viewing experience. To overcome these limitations, we propose the DenseSR, approaching the problem from a dense prediction perspective to emphasize restoration quality. This framework uniquely synergizes two key strategies: (1) deep scene understanding guided by geometric-semantic priors to resolve ambiguity and implicitly localize shadows, and (2) high-fidelity restoration via a novel Dense Fusion Block (DFB) in the decoder. The DFB employs adaptive component processing-using an Adaptive Content Smoothing Module (ACSM) for consistent appearance and a Texture-Boundary Recuperation Module (TBRM) for fine textures and sharp boundaries-thereby directly tackling the inconsistent restoration and blurring issues. These purposefully processed components are effectively fused, yielding an optimized feature representation preserving both consistency and fidelity. Extensive experimental results demonstrate the merits of our approach over existing methods. Our code can be available on https://github$.$com/VanLinLin/DenseSR

Ching-Heng Cheng, Chih-Chung Hsu

Remote sensing change detection (RSCD) aims to identify surface changes from co-registered bi-temporal images. However, many deep learning-based RSCD methods rely solely on change-map annotations and underuse the semantic information in non-changing regions, which limits robustness under illumination variation, off-nadir views, and scarce labels. This article introduces ChangeDINO, an end-to-end multiscale Siamese framework for optical building change detection. The model fuses a lightweight backbone stream with features transferred from a frozen DINOv3, yielding semantic- and context-rich pyramids even on small datasets. A spatial-spectral differential transformer decoder then exploits multi-scale absolute differences as change priors to highlight true building changes and suppress irrelevant responses. Finally, a learnable morphology module refines the upsampled logits to recover clean boundaries. Experiments on four public benchmarks show that ChangeDINO consistently outperforms recent state-of-the-art methods in IoU and F1, and ablation studies confirm the effectiveness of each component. The source code is available at https://github.com/chingheng0808/ChangeDINO.

Ching-Heng Cheng, Jen-Wei Lee, Chia-Ming Lee et al.

Underwater Image Enhancement (UIE) aims to restore visibility and correct color distortions caused by wavelength-dependent absorption and scattering. Recent hybrid approaches, which couple domain priors with modern deep neural architectures, have achieved strong performance but incur high computational cost, limiting their practicality in real-time scenarios. In this work, we propose WWE-UIE, a compact and efficient enhancement network that integrates three interpretable priors. First, adaptive white balance alleviates the strong wavelength-dependent color attenuation, particularly the dominance of blue-green tones. Second, a wavelet-based enhancement block (WEB) performs multi-band decomposition, enabling the network to capture both global structures and fine textures, which are critical for underwater restoration. Third, a gradient-aware module (SGFB) leverages Sobel operators with learnable gating to explicitly preserve edge structures degraded by scattering. Extensive experiments on benchmark datasets demonstrate that WWE-UIE achieves competitive restoration quality with substantially fewer parameters and FLOPs, enabling real-time inference on resource-limited platforms. Ablation studies and visualizations further validate the contribution of each component. The source code is available at https://github.com/chingheng0808/WWE-UIE.

Chih-Chung Hsu, Shao-Ning Chen, Mei-Hsuan Wu et al.

As DeepFake video manipulation techniques escalate, posing profound threats, the urgent need to develop efficient detection strategies is underscored. However, one particular issue lies with facial images being mis-detected, often originating from degraded videos or adversarial attacks, leading to unexpected temporal artifacts that can undermine the efficacy of DeepFake video detection techniques. This paper introduces a novel method for robust DeepFake video detection, harnessing the power of the proposed Graph-Regularized Attentive Convolutional Entanglement (GRACE) based on the graph convolutional network with graph Laplacian to address the aforementioned challenges. First, conventional Convolution Neural Networks are deployed to perform spatiotemporal features for the entire video. Then, the spatial and temporal features are mutually entangled by constructing a graph with sparse constraint, enforcing essential features of valid face images in the noisy face sequences remaining, thus augmenting stability and performance for DeepFake video detection. Furthermore, the Graph Laplacian prior is proposed in the graph convolutional network to remove the noise pattern in the feature space to further improve the performance. Comprehensive experiments are conducted to illustrate that our proposed method delivers state-of-the-art performance in DeepFake video detection under noisy face sequences. The source code is available at https://github.com/ming053l/GRACE.

Chih-Chung Hsu, Chih-Chien Ni, Chia-Ming Lee et al.

Hyperspectral imaging, capturing detailed spectral information for each pixel, is pivotal in diverse scientific and industrial applications. Yet, the acquisition of high-resolution (HR) hyperspectral images (HSIs) often needs to be addressed due to the hardware limitations of existing imaging systems. A prevalent workaround involves capturing both a high-resolution multispectral image (HR-MSI) and a low-resolution (LR) HSI, subsequently fusing them to yield the desired HR-HSI. Although deep learning-based methods have shown promising in HR-MSI/LR-HSI fusion and LR-HSI super-resolution (SR), their substantial model complexities hinder deployment on resource-constrained imaging devices. This paper introduces a novel knowledge distillation (KD) framework for HR-MSI/LR-HSI fusion to achieve SR of LR-HSI. Our KD framework integrates the proposed Cross-Layer Residual Aggregation (CLRA) block to enhance efficiency for constructing Dual Two-Streamed (DTS) network structure, designed to extract joint and distinct features from LR-HSI and HR-MSI simultaneously. To fully exploit the spatial and spectral feature representations of LR-HSI and HR-MSI, we propose a novel Cross Self-Attention (CSA) fusion module to adaptively fuse those features to improve the spatial and spectral quality of the reconstructed HR-HSI. Finally, the proposed KD-based joint loss function is employed to co-train the teacher and student networks. Our experimental results demonstrate that the student model not only achieves comparable or superior LR-HSI SR performance but also significantly reduces the model-size and computational requirements. This marks a substantial advancement over existing state-of-the-art methods. The source code is available at https://github.com/ming053l/CSAKD.

Chia-Ming Lee, Yu-Fan Lin, Yu-Hao Ho et al.

Hyperspectral image (HSI) fusion addresses the challenge of reconstructing High-Resolution HSIs (HR-HSIs) from High-Resolution Multispectral images (HR-MSIs) and Low-Resolution HSIs (LR-HSIs), a critical task given the high costs and hardware limitations associated with acquiring high-quality HSIs. While existing methods leverage spatial and spectral relationships, they often suffer from limited receptive fields and insufficient feature utilization, leading to suboptimal performance. Furthermore, the scarcity of high-quality HSI data highlights the importance of efficient data utilization to maximize reconstruction quality. To address these issues, we propose HyFusion, a novel Dual-Coupled Network (DCN) framework designed to enhance cross-domain feature extraction and enable effective feature map reusing. The framework first processes HR-MSI and LR-HSI inputs through specialized subnetworks that mutually enhance each other during feature extraction, preserving complementary spatial and spectral details. At its core, HyFusion utilizes an Enhanced Reception Field Block (ERFB), which combines shifting-window attention and dense connections to expand the receptive field, effectively capturing long-range dependencies while minimizing information loss. Extensive experiments demonstrate that HyFusion achieves state-of-the-art performance in HR-MSI/LR-HSI fusion, significantly improving reconstruction quality while maintaining a compact model size and computational efficiency. By integrating enhanced receptive fields and feature map reusing into a coupled network architecture, HyFusion provides a practical and effective solution for HSI fusion in resource-constrained scenarios, setting a new benchmark in hyperspectral imaging. Our code will be publicly available.

Yu-Fan Lin, Bo-Cheng Qiu, Chia-Ming Lee et al.

Accurate detection and segmentation of gastrointestinal bleeding are critical for diagnosing diseases such as peptic ulcers and colorectal cancer. This study proposes a two-stage framework that decouples classification and grounding to address the inherent challenges posed by traditional Multi-Task Learning models, which jointly optimizes classification and segmentation. Our approach separates these tasks to achieve targeted optimization for each. The model first classifies images as bleeding or non-bleeding, thereby isolating subsequent grounding from inter-task interference and label heterogeneity. To further enhance performance, we incorporate Stochastic Weight Averaging and Test-Time Augmentation, which improve model robustness against domain shifts and annotation inconsistencies. Our method is validated on the Auto-WCEBleedGen Challenge V2 Challenge dataset and achieving second place. Experimental results demonstrate significant improvements in classification accuracy and segmentation precision, especially on sequential datasets with consistent visual patterns. This study highlights the practical benefits of a two-stage strategy for medical image analysis and sets a new standard for GI bleeding detection and segmentation. Our code is publicly available at this GitHub repository.

Zheng Chen, Zongwei Wu, Eduard Zamfir et al.

This paper reviews the NTIRE 2024 challenge on image super-resolution ($\times$4), highlighting the solutions proposed and the outcomes obtained. The challenge involves generating corresponding high-resolution (HR) images, magnified by a factor of four, from low-resolution (LR) inputs using prior information. The LR images originate from bicubic downsampling degradation. The aim of the challenge is to obtain designs/solutions with the most advanced SR performance, with no constraints on computational resources (e.g., model size and FLOPs) or training data. The track of this challenge assesses performance with the PSNR metric on the DIV2K testing dataset. The competition attracted 199 registrants, with 20 teams submitting valid entries. This collective endeavour not only pushes the boundaries of performance in single-image SR but also offers a comprehensive overview of current trends in this field.

Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou et al.

This work examines the findings of the NTIRE 2025 Shadow Removal Challenge. A total of 306 participants have registered, with 17 teams successfully submitting their solutions during the final evaluation phase. Following the last two editions, this challenge had two evaluation tracks: one focusing on reconstruction fidelity and the other on visual perception through a user study. Both tracks were evaluated with images from the WSRD+ dataset, simulating interactions between self- and cast-shadows with a large number of diverse objects, textures, and materials.

Chih-Chung Hsu, Chih-Yu Jian, Eng-Shen Tu et al.

This paper addresses the challenges associated with hyperspectral image (HSI) reconstruction from miniaturized satellites, which often suffer from stripe effects and are computationally resource-limited. We propose a Real-Time Compressed Sensing (RTCS) network designed to be lightweight and require only relatively few training samples for efficient and robust HSI reconstruction in the presence of the stripe effect and under noisy transmission conditions. The RTCS network features a simplified architecture that reduces the required training samples and allows for easy implementation on integer-8-based encoders, facilitating rapid compressed sensing for stripe-like HSI, which exactly matches the moderate design of miniaturized satellites on push broom scanning mechanism. This contrasts optimization-based models that demand high-precision floating-point operations, making them difficult to deploy on edge devices. Our encoder employs an integer-8-compatible linear projection for stripe-like HSI data transmission, ensuring real-time compressed sensing. Furthermore, based on the novel two-streamed architecture, an efficient HSI restoration decoder is proposed for the receiver side, allowing for edge-device reconstruction without needing a sophisticated central server. This is particularly crucial as an increasing number of miniaturized satellites necessitates significant computing resources on the ground station. Extensive experiments validate the superior performance of our approach, offering new and vital capabilities for existing miniaturized satellite systems.

Benjamin Kiefer, Lojze Žust, Jon Muhovič et al.

The 3rd Workshop on Maritime Computer Vision (MaCVi) 2025 addresses maritime computer vision for Unmanned Surface Vehicles (USV) and underwater. This report offers a comprehensive overview of the findings from the challenges. We provide both statistical and qualitative analyses, evaluating trends from over 700 submissions. All datasets, evaluation code, and the leaderboard are available to the public at https://macvi.org/workshop/macvi25.

Chih-Chung Hsu, Chia-Ming Lee, Yang Fan Chiang et al.

This study explores the use of deep learning techniques for analyzing lung Computed Tomography (CT) images. Classic deep learning approaches face challenges with varying slice counts and resolutions in CT images, a diversity arising from the utilization of assorted scanning equipment. Typically, predictions are made on single slices which are then combined for a comprehensive outcome. Yet, this method does not incorporate learning features specific to each slice, leading to a compromise in effectiveness. To address these challenges, we propose an advanced Spatial-Slice Feature Learning (SSFL++) framework specifically tailored for CT scans. It aims to filter out out-of-distribution (OOD) data within the entire CT scan, allowing us to select essential spatial-slice features for analysis by reducing data redundancy by 70\%. Additionally, we introduce a Kernel-Density-based slice Sampling (KDS) method to enhance stability during training and inference phases, thereby accelerating convergence and enhancing overall performance. Remarkably, our experiments reveal that our model achieves promising results with a simple EfficientNet-2D (E2D) model. The effectiveness of our approach is confirmed on the COVID-19-CT-DB datasets provided by the DEF-AI-MIA workshop.

Jiun-Man Chen, Yu-Hsuan Chao, Yu-Jie Wang et al.

Transformer-based models have gained widespread popularity in both the computer vision (CV) and natural language processing (NLP) fields. However, significant challenges arise during post-training linear quantization, leading to noticeable reductions in inference accuracy. Our study focuses on uncovering the underlying causes of these accuracy drops and proposing a quantization-friendly fine-tuning method, \textbf{QuantTune}. Firstly, our analysis revealed that, on average, 65\% of quantization errors result from the precision loss incurred by the dynamic range amplification effect of outliers across the target Transformer-based models. Secondly, \textbf{QuantTune} adjusts weights based on the deviation of outlier activations and effectively constrains the dynamic ranges of the problematic activations. As a result, it successfully mitigates the negative impact of outliers on the inference accuracy of quantized models. Lastly, \textbf{QuantTune} can be seamlessly integrated into the back-propagation pass in the fine-tuning process without requiring extra complexity in inference software and hardware design. Our approach showcases significant improvements in post-training quantization across a range of Transformer-based models, including ViT, Bert-base, and OPT. QuantTune reduces accuracy drops by 12.09\% at 8-bit quantization and 33.8\% at 7-bit compared to top calibration methods, outperforming state-of-the-art solutions by over 18.84\% across ViT models.

Chia-Ming Lee, Yu-Fan Lin, Li-Wei Kang et al.

High-resolution hyperspectral imaging plays a crucial role in various remote sensing applications, yet its acquisition often faces fundamental limitations due to hardware constraints. This paper introduces S$^{3}$RNet, a novel framework for hyperspectral image pansharpening that effectively combines low-resolution hyperspectral images (LRHSI) with high-resolution multispectral images (HRMSI) through sparse spatial-spectral representation. The core of S$^{3}$RNet is the Multi-Branch Fusion Network (MBFN), which employs parallel branches to capture complementary features at different spatial and spectral scales. Unlike traditional approaches that treat all features equally, our Spatial-Spectral Attention Weight Block (SSAWB) dynamically adjusts feature weights to maintain sparse representation while suppressing noise and redundancy. To enhance feature propagation, we incorporate the Dense Feature Aggregation Block (DFAB), which efficiently aggregates inputted features through dense connectivity patterns. This integrated design enables S$^{3}$RNet to selectively emphasize the most informative features from differnt scale while maintaining computational efficiency. Comprehensive experiments demonstrate that S$^{3}$RNet achieves state-of-the-art performance across multiple evaluation metrics, showing particular strength in maintaining high reconstruction quality even under challenging noise conditions. The code will be made publicly available.

Chih-Chung Hsu, Chia-Ming Lee

Instance segmentation, a cornerstone task in computer vision, has wide-ranging applications in diverse industries. The advent of deep learning and artificial intelligence has underscored the criticality of training effective models, particularly in data-scarce scenarios - a concern that resonates in both academic and industrial circles. A significant impediment in this domain is the resource-intensive nature of procuring high-quality, annotated data for instance segmentation, a hurdle that amplifies the challenge of developing robust models under resource constraints. In this context, the strategic integration of a visual prior into the training dataset emerges as a potential solution to enhance congruity with the testing data distribution, consequently reducing the dependency on computational resources and the need for highly complex models. However, effectively embedding a visual prior into the learning process remains a complex endeavor. Addressing this challenge, we introduce the MISS (Memory-efficient Instance Segmentation System) framework. MISS leverages visual inductive prior flow propagation, integrating intrinsic prior knowledge from the Synergy-basketball dataset at various stages: data preprocessing, augmentation, training, and inference. Our empirical evaluations underscore the efficacy of MISS, demonstrating commendable performance in scenarios characterized by limited data availability and memory constraints.

Chih-Chung Hsu, Chia-Ming Lee, Ming-Shyen Wu

Instance segmentation is a fundamental task in computer vision with broad applications across various industries. In recent years, with the proliferation of deep learning and artificial intelligence applications, how to train effective models with limited data has become a pressing issue for both academia and industry. In the Visual Inductive Priors challenge (VIPriors2023), participants must train a model capable of precisely locating individuals on a basketball court, all while working with limited data and without the use of transfer learning or pre-trained models. We propose Memory effIciency inStance Segmentation framework based on visual inductive prior flow propagation that effectively incorporates inherent prior information from the dataset into both the data preprocessing and data augmentation stages, as well as the inference phase. Our team (ACVLAB) experiments demonstrate that our model achieves promising performance (0.509 AP@0.50:0.95) even under limited data and memory constraints.

Chih-Chung Hsu, Chia-Ming Lee, Chun-Hung Sun et al.

Automatic optical inspection (AOI) plays a pivotal role in the manufacturing process, predominantly leveraging high-resolution imaging instruments for scanning purposes. It detects anomalies by analyzing image textures or patterns, making it an essential tool in industrial manufacturing and quality control. Despite its importance, the deployment of models for AOI often faces challenges. These include limited sample sizes, which hinder effective feature learning, variations among source domains, and sensitivities to changes in lighting and camera positions during imaging. These factors collectively compromise the accuracy of model predictions. Traditional AOI often fails to capitalize on the rich mechanism-parameter information from machines or inside images, including statistical parameters, which typically benefit AOI classification. To address this, we introduce an external modality-guided data mining framework, primarily rooted in optical character recognition (OCR), to extract statistical features from images as a second modality to enhance performance, termed OANet (Ocr-Aoi-Net). A key aspect of our approach is the alignment of external modality features, extracted using a single modality-aware model, with image features encoded by a convolutional neural network. This synergy enables a more refined fusion of semantic representations from different modalities. We further introduce feature refinement and a gating function in our OANet to optimize the combination of these features, enhancing inference and decision-making capabilities. Experimental outcomes show that our methodology considerably boosts the recall rate of the defect detection model and maintains high robustness even in challenging scenarios.

Kailash A. Hambarde, Hugo Proença, Md Rashidunnabi et al.

Person re-identification (ReID) across aerial and ground views at extreme far distances introduces a distinct operating regime where severe resolution degradation, extreme viewpoint changes, unstable motion cues, and clothing variation jointly undermine the appearance-based assumptions of existing ReID systems. To study this regime, we introduce VReID-XFD, a video-based benchmark and community challenge for extreme far-distance (XFD) aerial-to-ground person re-identification. VReID-XFD is derived from the DetReIDX dataset and comprises 371 identities, 11,288 tracklets, and 11.75 million frames, captured across altitudes from 5.8 m to 120 m, viewing angles from oblique (30 degrees) to nadir (90 degrees), and horizontal distances up to 120 m. The benchmark supports aerial-to-aerial, aerial-to-ground, and ground-to-aerial evaluation under strict identity-disjoint splits, with rich physical metadata. The VReID-XFD-25 Challenge attracted 10 teams with hundreds of submissions. Systematic analysis reveals monotonic performance degradation with altitude and distance, a universal disadvantage of nadir views, and a trade-off between peak performance and robustness. Even the best-performing SAS-PReID method achieves only 43.93 percent mAP in the aerial-to-ground setting. The dataset, annotations, and official evaluation protocols are publicly available at https://www.it.ubi.pt/DetReIDX/ .

Chia-Ming Lee, Bo-Cheng Qiu, Ting-Yao Chen et al.

We present our solution for the Multi-Source COVID-19 Detection Challenge, which classifies chest CT scans from four distinct medical centers. To address multi-source variability, we employ the Spatial-Slice Feature Learning (SSFL) framework with Kernel-Density-based Slice Sampling (KDS). Our preprocessing pipeline combines lung region extraction, quality control, and adaptive slice sampling to select eight representative slices per scan. We compare EfficientNet and Swin Transformer architectures on the validation set. The EfficientNet model achieves an F1-score of 94.68%, compared to the Swin Transformer's 93.34%. The results demonstrate the effectiveness of our KDS-based pipeline on multi-source data and highlight the importance of dataset balance in multi-institutional medical imaging evaluation.

Chih-Chung Hsu, I-Hsuan Wu, Wen-Hai Tseng et al.

This report presents our semantic segmentation framework developed by team ACVLAB for the ICRA 2025 GOOSE 2D Semantic Segmentation Challenge, which focuses on parsing outdoor scenes into nine semantic categories under real-world conditions. Our method integrates a Swin Transformer backbone enhanced with Rotary Position Embedding (RoPE) for improved spatial generalization, alongside a Color Shift Estimation-and-Correction module designed to compensate for illumination inconsistencies in natural environments. To further improve training stability, we adopt a quantile-based denoising strategy that downweights the top 2.5\% of highest-error pixels, treating them as noise and suppressing their influence during optimization. Evaluated on the official GOOSE test set, our approach achieved a mean Intersection over Union (mIoU) of 0.848, demonstrating the effectiveness of combining color correction, positional encoding, and error-aware denoising in robust semantic segmentation.

Chih-Chung Hsu, Chia-Ming Lee, Chun-Hung Sun et al.

Traditional defect classification approaches are facing with two barriers. (1) Insufficient training data and unstable data quality. Collecting sufficient defective sample is expensive and time-costing, consequently leading to dataset variance. It introduces the difficulty on recognition and learning. (2) Over-dependence on visual modality. When the image pattern and texture is monotonic for all defect classes in a given dataset, the performance of conventional AOI system cannot be guaranteed. In scenarios where image quality is compromised due to mechanical failures or when defect information is inherently difficult to discern, the performance of deep models cannot be guaranteed. A main question is, "how to solve those two problems when they occur at the same time?" The feasible strategy is to explore another feature within dataset and combine an eminent vision-language model (VLM) and Large-Language model (LLM) with their astonishing zero-shot capability. In this work, we propose the special ASE dataset, including rich data description recorded on image, for defect classification, but the defect feature is uneasy to learn directly. Secondly, We present the prompting for VLM-LLM against defect classification with the proposed ASE dataset to activate extra-modality feature from images to enhance performance. Then, We design the novel progressive feature alignment (PFA) block to refine image-text feature to alleviate the difficulty of alignment under few-shot scenario. Finally, the proposed Cross-modality attention fusion (CMAF) module can effectively fuse different modality feature. Experiment results have demonstrated our method's effectiveness over several defect classification methods for the ASE dataset.

Huihui Fang, Fei Li, Huazhu Fu et al.

Age-related macular degeneration (AMD) is the leading cause of visual impairment among elderly in the world. Early detection of AMD is of great importance, as the vision loss caused by this disease is irreversible and permanent. Color fundus photography is the most cost-effective imaging modality to screen for retinal disorders. Cutting edge deep learning based algorithms have been recently developed for automatically detecting AMD from fundus images. However, there are still lack of a comprehensive annotated dataset and standard evaluation benchmarks. To deal with this issue, we set up the Automatic Detection challenge on Age-related Macular degeneration (ADAM), which was held as a satellite event of the ISBI 2020 conference. The ADAM challenge consisted of four tasks which cover the main aspects of detecting and characterizing AMD from fundus images, including detection of AMD, detection and segmentation of optic disc, localization of fovea, and detection and segmentation of lesions. As part of the challenge, we have released a comprehensive dataset of 1200 fundus images with AMD diagnostic labels, pixel-wise segmentation masks for both optic disc and AMD-related lesions (drusen, exudates, hemorrhages and scars, among others), as well as the coordinates corresponding to the location of the macular fovea. A uniform evaluation framework has been built to make a fair comparison of different models using this dataset. During the challenge, 610 results were submitted for online evaluation, with 11 teams finally participating in the onsite challenge. This paper introduces the challenge, the dataset and the evaluation methods, as well as summarizes the participating methods and analyzes their results for each task. In particular, we observed that the ensembling strategy and the incorporation of clinical domain knowledge were the key to improve the performance of the deep learning models.

Chih-Chung Hsu, Guan-Lin Chen, Mei-Hsuan Wu

With the massive damage in the world caused by Coronavirus Disease 2019 SARS-CoV-2 (COVID-19), many related research topics have been proposed in the past two years. The Chest Computed Tomography (CT) scans are the most valuable materials to diagnose the COVID-19 symptoms. However, most schemes for COVID-19 classification of Chest CT scan is based on a single-slice level, implying that the most critical CT slice should be selected from the original CT scan volume manually. We simultaneously propose 2-D and 3-D models to predict the COVID-19 of CT scan to tickle this issue. In our 2-D model, we introduce the Deep Wilcoxon signed-rank test (DWCC) to determine the importance of each slice of a CT scan to overcome the issue mentioned previously. Furthermore, a Convolutional CT scan-Aware Transformer (CCAT) is proposed to discover the context of the slices fully. The frame-level feature is extracted from each CT slice based on any backbone network and followed by feeding the features to our within-slice-Transformer (WST) to discover the context information in the pixel dimension. The proposed Between-Slice-Transformer (BST) is used to aggregate the extracted spatial-context features of every CT slice. A simple classifier is then used to judge whether the Spatio-temporal features are COVID-19 or non-COVID-19. The extensive experiments demonstrated that the proposed CCAT and DWCC significantly outperform the state-of-the-art methods.

Andrey Ignatov, Cheng-Ming Chiang, Hsien-Kai Kuo et al.

As the quality of mobile cameras starts to play a crucial role in modern smartphones, more and more attention is now being paid to ISP algorithms used to improve various perceptual aspects of mobile photos. In this Mobile AI challenge, the target was to develop an end-to-end deep learning-based image signal processing (ISP) pipeline that can replace classical hand-crafted ISPs and achieve nearly real-time performance on smartphone NPUs. For this, the participants were provided with a novel learned ISP dataset consisting of RAW-RGB image pairs captured with the Sony IMX586 Quad Bayer mobile sensor and a professional 102-megapixel medium format camera. The runtime of all models was evaluated on the MediaTek Dimensity 1000+ platform with a dedicated AI processing unit capable of accelerating both floating-point and quantized neural networks. The proposed solutions are fully compatible with the above NPU and are capable of processing Full HD photos under 60-100 milliseconds while achieving high fidelity results. A detailed description of all models developed in this challenge is provided in this paper.

Chih-Chung Hsu, Hsin-Ti Ma

Semantic segmentation is one of the most attractive research fields in computer vision. In the VIPriors challenge, only very limited numbers of training samples are allowed, leading to that the current state-of-the-art and deep learning-based semantic segmentation techniques are hard to train well. To overcome this shortcoming, therefore, we propose edge-preserving guidance to obtain the extra prior information, to avoid the overfitting under small-scale training dataset. First, a two-channeled convolutional layer is concatenated to the last layer of the conventional semantic segmentation network. Then, an edge map is calculated from the ground truth by Sobel operation and followed by concatenating a hard-thresholding operation to indicate whether the pixel is the edge or not. Then, the two-dimensional cross-entropy loss is adopted to calculate the loss between the predicted edge map and its ground truth, termed as an edge-preserving loss. In this way, the continuity of boundaries between different instances can be forced by the proposed edge-preserving loss. Experiments demonstrate that the proposed method can achieve excellent performance under small-scale training set, compared to state-of-the-art semantic segmentation techniques.

Chih-Chung Hsu, Chia-Hsiang Lin

Deep learning-based single image super-resolution enables very fast and high-visual-quality reconstruction. Recently, an enhanced super-resolution based on generative adversarial network (ESRGAN) has achieved excellent performance in terms of both qualitative and quantitative quality of the reconstructed high-resolution image. In this paper, we propose to add one more shortcut between two dense-blocks, as well as add shortcut between two convolution layers inside a dense-block. With this simple strategy of adding more shortcuts in the proposed network, it enables a faster learning process as the gradient information can be back-propagated more easily. Based on the improved ESRGAN, the dual reconstruction is proposed to learn different aspects of the super-resolved image for judiciously enhancing the quality of the reconstructed image. In practice, the super-resolution model is pre-trained solely based on pixel distance, followed by fine-tuning the parameters in the model based on adversarial loss and perceptual loss. Finally, we fuse two different models by weighted-summing their parameters to obtain the final super-resolution model. Experimental results demonstrated that the proposed method achieves excellent performance in the real-world image super-resolution challenge. We have also verified that the proposed dual reconstruction does further improve the quality of the reconstructed image in terms of both PSNR and SSIM.

Andreas Lugmayr, Martin Danelljan, Radu Timofte et al.

This paper reviews the AIM 2019 challenge on real world super-resolution. It focuses on the participating methods and final results. The challenge addresses the real world setting, where paired true high and low-resolution images are unavailable. For training, only one set of source input images is therefore provided in the challenge. In Track 1: Source Domain the aim is to super-resolve such images while preserving the low level image characteristics of the source input domain. In Track 2: Target Domain a set of high-quality images is also provided for training, that defines the output domain and desired quality of the super-resolved images. To allow for quantitative evaluation, the source input images in both tracks are constructed using artificial, but realistic, image degradations. The challenge is the first of its kind, aiming to advance the state-of-the-art and provide a standard benchmark for this newly emerging task. In total 7 teams competed in the final testing phase, demonstrating new and innovative solutions to the problem.

Chih-Chung Hsu, Chia-Yen Lee, Yi-Xiu Zhuang

Although Generative Adversarial Network (GAN) can be used to generate the realistic image, improper use of these technologies brings hidden concerns. For example, GAN can be used to generate a tampered video for specific people and inappropriate events, creating images that are detrimental to a particular person, and may even affect that personal safety. In this paper, we will develop a deep forgery discriminator (DeepFD) to efficiently and effectively detect the computer-generated images. Directly learning a binary classifier is relatively tricky since it is hard to find the common discriminative features for judging the fake images generated from different GANs. To address this shortcoming, we adopt contrastive loss in seeking the typical features of the synthesized images generated by different GANs and follow by concatenating a classifier to detect such computer-generated images. Experimental results demonstrate that the proposed DeepFD successfully detected 94.7% fake images generated by several state-of-the-art GANs.

Chih-Chung Hsu, Chia-Yen Lee, Ting-Xuan Liao et al.

In this study, we propose a novel iterative refinement approach to predict the popularity score of the social media meta-data effectively. With the rapid growth of the social media on the Internet, how to adequately forecast the view count or popularity becomes more important. Conventionally, the ensemble approach such as random forest regression achieves high and stable performance on various prediction tasks. However, most of the regression methods may not precisely predict the extreme high or low values. To address this issue, we first predict the initial popularity score and retrieve their residues. In order to correctly compensate those extreme values, we adopt an ensemble regressor to compensate the residues to further improve the prediction performance. Comprehensive experiments are conducted to demonstrate the proposed iterative refinement approach outperforms the state-of-the-art regression approach.

Chih-Chung Hsu, Chia-Wen Lin, Weng-Tai Su et al.

Despite generative adversarial networks (GANs) can hallucinate photo-realistic high-resolution (HR) faces from low-resolution (LR) faces, they cannot guarantee preserving the identities of hallucinated HR faces, making the HR faces poorly recognizable. To address this problem, we propose a Siamese GAN (SiGAN) to reconstruct HR faces that visually resemble their corresponding identities. On top of a Siamese network, the proposed SiGAN consists of a pair of two identical generators and one discriminator. We incorporate reconstruction error and identity label information in the loss function of SiGAN in a pairwise manner. By iteratively optimizing the loss functions of the generator pair and discriminator of SiGAN, we cannot only achieve photo-realistic face reconstruction, but also ensures the reconstructed information is useful for identity recognition. Experimental results demonstrate that SiGAN significantly outperforms existing face hallucination GANs in objective face verification performance, while achieving photo-realistic reconstruction. Moreover, for input LR faces from unknown identities who are not included in training, SiGAN can still do a good job.

Chih-Chung Hsu, Chia-Wen Lin

Given a large unlabeled set of images, how to efficiently and effectively group them into clusters based on extracted visual representations remains a challenging problem. To address this problem, we propose a convolutional neural network (CNN) to jointly solve clustering and representation learning in an iterative manner. In the proposed method, given an input image set, we first randomly pick k samples and extract their features as initial cluster centroids using the proposed CNN with an initial model pre-trained from the ImageNet dataset. Mini-batch k-means is then performed to assign cluster labels to individual input samples for a mini-batch of images randomly sampled from the input image set until all images are processed. Subsequently, the proposed CNN simultaneously updates the parameters of the proposed CNN and the centroids of image clusters iteratively based on stochastic gradient descent. We also proposed a feature drift compensation scheme to mitigate the drift error caused by feature mismatch in representation learning. Experimental results demonstrate the proposed method outperforms start-of-the-art clustering schemes in terms of accuracy and storage complexity on large-scale image sets containing millions of images.