Jun-Wei Hsieh

Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen et al.

Despite recent progress in Multiple Object Tracking (MOT), several obstacles such as occlusions, similar objects, and complex scenes remain an open challenge. Meanwhile, a systematic study of the cost-performance tradeoff for the popular tracking-by-detection paradigm is still lacking. This paper introduces SMILEtrack, an innovative object tracker that effectively addresses these challenges by integrating an efficient object detector with a Siamese network-based Similarity Learning Module (SLM). The technical contributions of SMILETrack are twofold. First, we propose an SLM that calculates the appearance similarity between two objects, overcoming the limitations of feature descriptors in Separate Detection and Embedding (SDE) models. The SLM incorporates a Patch Self-Attention (PSA) block inspired by the vision Transformer, which generates reliable features for accurate similarity matching. Second, we develop a Similarity Matching Cascade (SMC) module with a novel GATE function for robust object matching across consecutive video frames, further enhancing MOT performance. Together, these innovations help SMILETrack achieve an improved trade-off between the cost ({\em e.g.}, running speed) and performance (e.g., tracking accuracy) over several existing state-of-the-art benchmarks, including the popular BYTETrack method. SMILETrack outperforms BYTETrack by 0.4-0.8 MOTA and 2.1-2.2 HOTA points on MOT17 and MOT20 datasets. Code is available at https://github.com/pingyang1117/SMILEtrack_Official

Chao-Peng Chen, Jun-Wei Hsieh, Ping-Yang Chen et al.

Change detection (CD) aims to find the difference between two images at different times and outputs a change map to represent whether the region has changed or not. To achieve a better result in generating the change map, many State-of-The-Art (SoTA) methods design a deep learning model that has a powerful discriminative ability. However, these methods still get lower performance because they ignore spatial information and scaling changes between objects, giving rise to blurry or wrong boundaries. In addition to these, they also neglect the interactive information of two different images. To alleviate these problems, we propose our network, the Scale and Relation-Aware Siamese Network (SARAS-Net) to deal with this issue. In this paper, three modules are proposed that include relation-aware, scale-aware, and cross-transformer to tackle the problem of scene change detection more effectively. To verify our model, we tested three public datasets, including LEVIR-CD, WHU-CD, and DSFIN, and obtained SoTA accuracy. Our code is available at https://github.com/f64051041/SARAS-Net.

Jun-Wei Hsieh, Meng-Yu Kao, Ghufron Wahyu Kurniawan et al.

YOLO-series and DETR-based detectors struggle with tiny-object detection. YOLO-style models benefit from efficient dense prediction, but their large-stride backbones may suppress tiny instances in deep feature maps and make grid assignment ambiguous. DETR-based models remove hand-crafted post-processing through set prediction, yet they reason over coarse token grids, where tiny objects occupy only a few weak tokens and are easily overlooked during matching. To address these limitations, we propose TinyFormer, a unified YOLO--DETR hybrid real-time detector that combines ViT representations, NMS-free set prediction, and a YOLO-style pyramid neck for accurate small-object detection. TinyFormer introduces a Parallel Bi-fusion Module (PBM), which builds high-resolution shortcuts from shallow stages to the feature pyramid, preserving fine spatial details during multi-scale fusion. We further design a Spatial Semantic Adapter (SSA) to compensate for the spatial loss caused by coarse tokenization. SSA extracts high-resolution cues from early stages and injects them into transformer token embeddings, improving tiny-object localization without sacrificing the global modeling ability of DETR. Experiments on MS COCO show that TinyFormer consistently outperforms recent YOLO-series detectors and the strong DEIMv2 baseline. TinyFormer-X achieves 58.4% AP even without PBM, while adding PBM improves the overall AP to 58.5% and brings a 1.6% AP gain on small objects. With Objects365 pre-training, TinyFormer-X-PBM reaches 60.2% AP, surpassing RF-DETR and other Objects365-pretrained detectors with fewer parameters and lower computation. These results demonstrate that TinyFormer bridges dense YOLO-style feature fusion and DETR-style set prediction, providing a strong accuracy-efficiency trade-off for real-time tiny-object detection. Code is available at https://github.com/mmpmmpmmpjosh/TinyFormer.

Yu-Xiang Zeng, Jun-Wei Hsieh, Xin Li et al.

Detecting small scene text instances in the wild is particularly challenging, where the influence of irregular positions and nonideal lighting often leads to detection errors. We present MixNet, a hybrid architecture that combines the strengths of CNNs and Transformers, capable of accurately detecting small text from challenging natural scenes, regardless of the orientations, styles, and lighting conditions. MixNet incorporates two key modules: (1) the Feature Shuffle Network (FSNet) to serve as the backbone and (2) the Central Transformer Block (CTBlock) to exploit the 1D manifold constraint of the scene text. We first introduce a novel feature shuffling strategy in FSNet to facilitate the exchange of features across multiple scales, generating high-resolution features superior to popular ResNet and HRNet. The FSNet backbone has achieved significant improvements over many existing text detection methods, including PAN, DB, and FAST. Then we design a complementary CTBlock to leverage center line based features similar to the medial axis of text regions and show that it can outperform contour-based approaches in challenging cases when small scene texts appear closely. Extensive experimental results show that MixNet, which mixes FSNet with CTBlock, achieves state-of-the-art results on multiple scene text detection datasets.

Yu-Ming Zhang, Jun-Wei Hsieh, Chun-Chieh Lee et al.

FPN (Feature Pyramid Network) has become a basic component of most SoTA one stage object detectors. Many previous studies have repeatedly proved that FPN can caputre better multi-scale feature maps to more precisely describe objects if they are with different sizes. However, for most backbones such VGG, ResNet, or DenseNet, the feature maps at each layer are downsized to their quarters due to the pooling operation or convolutions with stride 2. The gap of down-scaling-by-2 is large and makes its FPN not fuse the features smoothly. This paper proposes a new SFPN (Synthetic Fusion Pyramid Network) arichtecture which creates various synthetic layers between layers of the original FPN to enhance the accuracy of light-weight CNN backones to extract objects' visual features more accurately. Finally, experiments prove the SFPN architecture outperforms either the large backbone VGG16, ResNet50 or light-weight backbones such as MobilenetV2 based on AP score.

Benjamin Kiefer, Jan Lukas Augustin, Jon Muhovič et al.

The 4th Workshop on Maritime Computer Vision (MaCVi) is organized as part of CVPR 2026. This edition features five benchmark challenges with emphasis on both predictive accuracy and embedded real-time feasibility. This report summarizes the MaCVi 2026 challenge setup, evaluation protocols, datasets, and benchmark tracks, and presents quantitative results, qualitative comparisons, and cross-challenge analyses of emerging method trends. We also include technical reports from top-performing teams to highlight practical design choices and lessons learned across the benchmark suite. Datasets, leaderboards, and challenge resources are available at https://macvi.org/workshop/cvpr26.

Novendra Setyawan, Chi-Chia Sun, Mao-Hsiu Hsu et al.

Lightweight face recognition is increasingly important for deployment on edge and mobile devices, where strict constraints on latency, memory, and energy consumption must be met alongside reliable accuracy. Although recent hybrid CNN-Transformer architectures have advanced global context modeling, striking an effective balance between recognition performance and computational efficiency remains an open challenge. In this work, we present FaceLiVTv2, an improved version of our FaceLiVT hybrid architecture designed for efficient global--local feature interaction in mobile face recognition. At its core is Lite MHLA, a lightweight global token interaction module that replaces the original multi-layer attention design with multi-head linear token projections and affine rescale transformations, reducing redundancy while preserving representational diversity across heads. We further integrate Lite MHLA into a unified RepMix block that coordinates local and global feature interactions and adopts global depthwise convolution for adaptive spatial aggregation in the embedding stage. Under our experimental setup, results on LFW, CA-LFW, CP-LFW, CFP-FP, AgeDB-30, and IJB show that FaceLiVTv2 consistently improves the accuracy-efficiency trade-off over existing lightweight methods. Notably, FaceLiVTv2 reduces mobile inference latency by 22% relative to FaceLiVTv1, achieves speedups of up to 30.8% over GhostFaceNets on mobile devices, and delivers 20-41% latency improvements over EdgeFace and KANFace across platforms while maintaining higher recognition accuracy. These results demonstrate that FaceLiVTv2 offers a practical and deployable solution for real-time face recognition. Code is available at https://github.com/novendrastywn/FaceLiVT.

Yi-Kuan Hsieh, Jun-Wei Hsieh, Yu-Chee Tseng et al.

We develop a Synthetic Fusion Pyramid Network (SPF-Net) with a scale-aware loss function design for accurate crowd counting. Existing crowd-counting methods assume that the training annotation points were accurate and thus ignore the fact that noisy annotations can lead to large model-learning bias and counting error, especially for counting highly dense crowds that appear far away. To the best of our knowledge, this work is the first to properly handle such noise at multiple scales in end-to-end loss design and thus push the crowd counting state-of-the-art. We model the noise of crowd annotation points as a Gaussian and derive the crowd probability density map from the input image. We then approximate the joint distribution of crowd density maps with the full covariance of multiple scales and derive a low-rank approximation for tractability and efficient implementation. The derived scale-aware loss function is used to train the SPF-Net. We show that it outperforms various loss functions on four public datasets: UCF-QNRF, UCF CC 50, NWPU and ShanghaiTech A-B datasets. The proposed SPF-Net can accurately predict the locations of people in the crowd, despite training on noisy training annotations.

Novendra Setyawan, Chi-Chia Sun, Mao-Hsiu Hsu et al.

The Vision Transformer (ViT) achieves remarkable accuracy across visual tasks but remains computationally expensive for edge deployment. This paper presents MicroViTv2, a lightweight Vision Transformer optimized for real-device efficiency. Built upon the original MicroViT, the proposed model is designed based on reparameterized design, specifically Reparameterized Patch Embedding (RepEmbed) and Reparameterized Depth-Wise convolution mixer (RepDW) for faster inference, and introduces the Single Depth-Wise Transposed Attention (SDTA) to capture long-range dependencies with minimal redundancy. Despite slightly higher FLOPs, MicroViTv2 improves accuracy up to 0.5% compared to its predecessor and surpassing MobileViTv2, EdgeNeXt, and EfficientViT while maintaining fast inference and high energy efficiency on Jetson AGX Orin. Experiments on ImageNet-1K and COCO demonstrate that hardware-aware design and structural re-parameterization are key to achieving high accuracy and low energy consumption, validating the need to evaluate efficiency beyond FLOPs. Code is available at https://github.com/novendrastywn/MicroViT.

Yi-Chun Wang, Jun-Wei Hsieh, Ming-Ching Chang

Current NAS-based semantic segmentation methods focus on accuracy improvements rather than light-weight design. In this paper, we proposed a two-stage framework to design our NAS-based RSPNet model for light-weight semantic segmentation. The first architecture search determines the inner cell structure, and the second architecture search considers exponentially growing paths to finalize the outer structure of the network. It was shown in the literature that the fusion of high- and low-resolution feature maps produces stronger representations. To find the expected macro structure without manual design, we adopt a new path-attention mechanism to efficiently search for suitable paths to fuse useful information for better segmentation. Our search for repeatable micro-structures from cells leads to a superior network architecture in semantic segmentation. In addition, we propose an RSP (recursive Stage Partial) architecture to search a light-weight design for NAS-based semantic segmentation. The proposed architecture is very efficient, simple, and effective that both the macro- and micro- structure searches can be completed in five days of computation on two V100 GPUs. The light-weight NAS architecture with only 1/4 parameter size of SoTA architectures can achieve SoTA performance on semantic segmentation on the Cityscapes dataset without using any backbones.

Yu-Ming Zhang, Jun-Wei Hsieh, Chun-Chieh Lee et al.

In the past decade, many architectures of convolution neural networks were designed by handcraft, such as Vgg16, ResNet, DenseNet, etc. They all achieve state-of-the-art level on different tasks in their time. However, it still relies on human intuition and experience, and it also takes so much time consumption for trial and error. Neural Architecture Search (NAS) focused on this issue. In recent works, the Neural Predictor has significantly improved with few training architectures as training samples. However, the sampling efficiency is already considerable. In this paper, our proposed Siamese-Predictor is inspired by past works of predictor-based NAS. It is constructed with the proposed Estimation Code, which is the prior knowledge about the training procedure. The proposed Siamese-Predictor gets significant benefits from this idea. This idea causes it to surpass the current SOTA predictor on NASBench-201. In order to explore the impact of the Estimation Code, we analyze the relationship between it and accuracy. We also propose the search space Tiny-NanoBench for lightweight CNN architecture. This well-designed search space is easier to find better architecture with few FLOPs than NASBench-201. In summary, the proposed Siamese-Predictor is a predictor-based NAS. It achieves the SOTA level, especially with limited computation budgets. It applied to the proposed Tiny-NanoBench can just use a few trained samples to find extremely lightweight CNN architecture.

Ying-Yu Chen, Jun-Wei Hsieh, Ming-Ching Chang

Few-Shot Learning (FSL) has attracted growing attention in computer vision due to its capability in model training without the need for excessive data. FSL is challenging because the training and testing categories (the base vs. novel sets) can be largely diversified. Conventional transfer-based solutions that aim to transfer knowledge learned from large labeled training sets to target testing sets are limited, as critical adverse impacts of the shift in task distribution are not adequately addressed. In this paper, we extend the solution of transfer-based methods by incorporating the concept of metric-learning and channel attention. To better exploit the feature representations extracted by the feature backbone, we propose Class-Specific Channel Attention (CSCA) module, which learns to highlight the discriminative channels in each class by assigning each class one CSCA weight vector. Unlike general attention modules designed to learn global-class features, the CSCA module aims to learn local and class-specific features with very effective computation. We evaluated the performance of the CSCA module on standard benchmarks including miniImagenet, Tiered-ImageNet, CIFAR-FS, and CUB-200-2011. Experiments are performed in inductive and in/cross-domain settings. We achieve new state-of-the-art results.

Chi-Chun Chao, Jun-Wei Hsieh, Bor-Shiun Wang

Traffic signal control is a challenging real-world problem aiming to minimize overall travel time by coordinating vehicle movements at road intersections. Existing traffic signal control systems in use still rely heavily on oversimplified information and rule-based methods. Specifically, the periodicity of green/red light alternations can be considered as a prior for better planning of each agent in policy optimization. To better learn such adaptive and predictive priors, traditional RL-based methods can only return a fixed length from predefined action pool with only local agents. If there is no cooperation between these agents, some agents often make conflicts to other agents and thus decrease the whole throughput. This paper proposes a cooperative, multi-objective architecture with age-decaying weights to better estimate multiple reward terms for traffic signal control optimization, which termed COoperative Multi-Objective Multi-Agent Deep Deterministic Policy Gradient (COMMA-DDPG). Two types of agents running to maximize rewards of different goals - one for local traffic optimization at each intersection and the other for global traffic waiting time optimization. The global agent is used to guide the local agents as a means for aiding faster learning but not used in the inference phase. We also provide an analysis of solution existence together with convergence proof for the proposed RL optimization. Evaluation is performed using real-world traffic data collected using traffic cameras from an Asian country. Our method can effectively reduce the total delayed time by 60\%. Results demonstrate its superiority when compared to SoTA methods.

Mas Nurul Achmadiah, Chi-Chia Sun, Wen-Kai Kuo et al.

Crowd counting remains challenging in variable-density scenes due to scale variations, occlusions, and the high computational cost of existing models. To address these issues, we propose RepSFNet (Reparameterized Single Fusion Network), a lightweight architecture designed for accurate and real-time crowd estimation. RepSFNet leverages a RepLK-ViT backbone with large reparameterized kernels for efficient multi-scale feature extraction. It further integrates a Feature Fusion module combining Atrous Spatial Pyramid Pooling (ASPP) and Context-Aware Network (CAN) to achieve robust, density-adaptive context modeling. A Concatenate Fusion module is employed to preserve spatial resolution and generate high-quality density maps. By avoiding attention mechanisms and multi-branch designs, RepSFNet significantly reduces parameters and computational complexity. The training objective combines Mean Squared Error and Optimal Transport loss to improve both count accuracy and spatial distribution alignment. Experiments conducted on ShanghaiTech, NWPU, and UCF-QNRF datasets demonstrate that RepSFNet achieves competitive accuracy while reducing inference latency by up to 34 percent compared to recent state-of-the-art methods, making it suitable for real-time and low-power edge computing applications.

Ping-Yang Chen, Ming-Ching Chang, Jun-Wei Hsieh et al.

This paper proposes the Parallel Residual Bi-Fusion Feature Pyramid Network (PRB-FPN) for fast and accurate single-shot object detection. Feature Pyramid (FP) is widely used in recent visual detection, however the top-down pathway of FP cannot preserve accurate localization due to pooling shifting. The advantage of FP is weakened as deeper backbones with more layers are used. In addition, it cannot keep up accurate detection of both small and large objects at the same time. To address these issues, we propose a new parallel FP structure with bi-directional (top-down and bottom-up) fusion and associated improvements to retain high-quality features for accurate localization. We provide the following design improvements: (1) A parallel bifusion FP structure with a bottom-up fusion module (BFM) to detect both small and large objects at once with high accuracy. (2) A concatenation and re-organization (CORE) module provides a bottom-up pathway for feature fusion, which leads to the bi-directional fusion FP that can recover lost information from lower-layer feature maps. (3) The CORE feature is further purified to retain richer contextual information. Such CORE purification in both top-down and bottom-up pathways can be finished in only a few iterations. (4) The adding of a residual design to CORE leads to a new Re-CORE module that enables easy training and integration with a wide range of deeper or lighter backbones. The proposed network achieves state-of-the-art performance on the UAVDT17 and MS COCO datasets. Code is available at https://github.com/pingyang1117/PRBNet_PyTorch.

Chien-Yao Wang, Hong-Yuan Mark Liao, I-Hau Yeh et al.

Neural networks have enabled state-of-the-art approaches to achieve incredible results on computer vision tasks such as object detection. However, such success greatly relies on costly computation resources, which hinders people with cheap devices from appreciating the advanced technology. In this paper, we propose Cross Stage Partial Network (CSPNet) to mitigate the problem that previous works require heavy inference computations from the network architecture perspective. We attribute the problem to the duplicate gradient information within network optimization. The proposed networks respect the variability of the gradients by integrating feature maps from the beginning and the end of a network stage, which, in our experiments, reduces computations by 20% with equivalent or even superior accuracy on the ImageNet dataset, and significantly outperforms state-of-the-art approaches in terms of AP50 on the MS COCO object detection dataset. The CSPNet is easy to implement and general enough to cope with architectures based on ResNet, ResNeXt, and DenseNet. Source code is at https://github.com/WongKinYiu/CrossStagePartialNetworks.

Yi-Kuan Hsieh, Jun-Wei Hsieh, Xin li et al.

State Space Models (SSMs) have emerged as efficient alternatives to attention for vision tasks, offering lineartime sequence processing with competitive accuracy. Vision SSMs, however, require serializing 2D images into 1D token sequences along a predefined scan order, a factor often overlooked. We show that scan order critically affects performance by altering spatial adjacency, fracturing object continuity, and amplifying degradation under geometric transformations such as rotation. We present Partial RIng Scan Mamba (PRISMamba), a rotation-robust traversal that partitions an image into concentric rings, performs order-agnostic aggregation within each ring, and propagates context across rings through a set of short radial SSMs. Efficiency is further improved via partial channel filtering, which routes only the most informative channels through the recurrent ring pathway while keeping the rest on a lightweight residual branch. On ImageNet-1K, PRISMamba achieves 84.5% Top-1 with 3.9G FLOPs and 3,054 img/s on A100, outperforming VMamba in both accuracy and throughput while requiring fewer FLOPs. It also maintains performance under rotation, whereas fixed-path scans drop by 1~2%. These results highlight scan-order design, together with channel filtering, as a crucial, underexplored factor for accuracy, efficiency, and rotation robustness in Vision SSMs. Code will be released upon acceptance.

Shuo Wang, David C. Anastasiu, Zheng Tang et al. · mit

The eighth AI City Challenge highlighted the convergence of computer vision and artificial intelligence in areas like retail, warehouse settings, and Intelligent Traffic Systems (ITS), presenting significant research opportunities. The 2024 edition featured five tracks, attracting unprecedented interest from 726 teams in 47 countries and regions. Track 1 dealt with multi-target multi-camera (MTMC) people tracking, highlighting significant enhancements in camera count, character number, 3D annotation, and camera matrices, alongside new rules for 3D tracking and online tracking algorithm encouragement. Track 2 introduced dense video captioning for traffic safety, focusing on pedestrian accidents using multi-camera feeds to improve insights for insurance and prevention. Track 3 required teams to classify driver actions in a naturalistic driving analysis. Track 4 explored fish-eye camera analytics using the FishEye8K dataset. Track 5 focused on motorcycle helmet rule violation detection. The challenge utilized two leaderboards to showcase methods, with participants setting new benchmarks, some surpassing existing state-of-the-art achievements.

Xian-Hong Huang, Hui-Kai Su, Chi-Chia Sun et al.

This paper introduces a cutting-edge approach to cross-modal interaction for tiny object detection by combining semantic-guided natural language processing with advanced visual recognition backbones. The proposed method integrates the BERT language model with the CNN-based Parallel Residual Bi-Fusion Feature Pyramid Network (PRB-FPN-Net), incorporating innovative backbone architectures such as ELAN, MSP, and CSP to optimize feature extraction and fusion. By employing lemmatization and fine-tuning techniques, the system aligns semantic cues from textual inputs with visual features, enhancing detection precision for small and complex objects. Experimental validation using the COCO and Objects365 datasets demonstrates that the model achieves superior performance. On the COCO2017 validation set, it attains a 52.6% average precision (AP), outperforming YOLO-World significantly while maintaining half the parameter consumption of Transformer-based models like GLIP. Several test on different of backbones such ELAN, MSP, and CSP further enable efficient handling of multi-scale objects, ensuring scalability and robustness in resource-constrained environments. This study underscores the potential of integrating natural language understanding with advanced backbone architectures, setting new benchmarks in object detection accuracy, efficiency, and adaptability to real-world challenges.

Novendra Setyawan, Chi-Chia Sun, Mao-Hsiu Hsu et al.

The Vision Transformer (ViT) has demonstrated state-of-the-art performance in various computer vision tasks, but its high computational demands make it impractical for edge devices with limited resources. This paper presents MicroViT, a lightweight Vision Transformer architecture optimized for edge devices by significantly reducing computational complexity while maintaining high accuracy. The core of MicroViT is the Efficient Single Head Attention (ESHA) mechanism, which utilizes group convolution to reduce feature redundancy and processes only a fraction of the channels, thus lowering the burden of the self-attention mechanism. MicroViT is designed using a multi-stage MetaFormer architecture, stacking multiple MicroViT encoders to enhance efficiency and performance. Comprehensive experiments on the ImageNet-1K and COCO datasets demonstrate that MicroViT achieves competitive accuracy while significantly improving 3.6 faster inference speed and reducing energy consumption with 40% higher efficiency than the MobileViT series, making it suitable for deployment in resource-constrained environments such as mobile and edge devices.

Zheng Tang, Shuo Wang, David C. Anastasiu et al.

The ninth AI City Challenge continues to advance real-world applications of computer vision and AI in transportation, industrial automation, and public safety. The 2025 edition featured four tracks and saw a 17% increase in participation, with 245 teams from 15 countries registered on the evaluation server. Public release of challenge datasets led to over 30,000 downloads to date. Track 1 focused on multi-class 3D multi-camera tracking, involving people, humanoids, autonomous mobile robots, and forklifts, using detailed calibration and 3D bounding box annotations. Track 2 tackled video question answering in traffic safety, with multi-camera incident understanding enriched by 3D gaze labels. Track 3 addressed fine-grained spatial reasoning in dynamic warehouse environments, requiring AI systems to interpret RGB-D inputs and answer spatial questions that combine perception, geometry, and language. Both Track 1 and Track 3 datasets were generated in NVIDIA Omniverse. Track 4 emphasized efficient road object detection from fisheye cameras, supporting lightweight, real-time deployment on edge devices. The evaluation framework enforced submission limits and used a partially held-out test set to ensure fair benchmarking. Final rankings were revealed after the competition concluded, fostering reproducibility and mitigating overfitting. Several teams achieved top-tier results, setting new benchmarks in multiple tasks.

Novendra Setyawan, Chi-Chia Sun, Mao-Hsiu Hsu et al.

This paper introduces FaceLiVT, a lightweight yet powerful face recognition model that integrates a hybrid Convolution Neural Network (CNN)-Transformer architecture with an innovative and lightweight Multi-Head Linear Attention (MHLA) mechanism. By combining MHLA alongside a reparameterized token mixer, FaceLiVT effectively reduces computational complexity while preserving competitive accuracy. Extensive evaluations on challenging benchmarks; including LFW, CFP-FP, AgeDB-30, IJB-B, and IJB-C; highlight its superior performance compared to state-of-the-art lightweight models. MHLA notably improves inference speed, allowing FaceLiVT to deliver high accuracy with lower latency on mobile devices. Specifically, FaceLiVT is 8.6 faster than EdgeFace, a recent hybrid CNN-Transformer model optimized for edge devices, and 21.2 faster than a pure ViT-Based model. With its balanced design, FaceLiVT offers an efficient and practical solution for real-time face recognition on resource-constrained platforms.

Yi-Kuan Hsieh, Jun-Wei Hsieh, Xin Li et al.

Vision Transformer (ViT) has achieved impressive results across various vision tasks, yet its high computational cost limits practical applications. Recent methods have aimed to reduce ViT's $O(n^2)$ complexity by pruning unimportant tokens. However, these techniques often sacrifice accuracy by independently pruning query (Q) and key (K) tokens, leading to performance degradation due to overlooked token interactions. To address this limitation, we introduce a novel {\bf Block-based Symmetric Pruning and Fusion} for efficient ViT (BSPF-ViT) that optimizes the pruning of Q/K tokens jointly. Unlike previous methods that consider only a single direction, our approach evaluates each token and its neighbors to decide which tokens to retain by taking token interaction into account. The retained tokens are compressed through a similarity fusion step, preserving key information while reducing computational costs. The shared weights of Q/K tokens create a symmetric attention matrix, allowing pruning only the upper triangular part for speed up. BSPF-ViT consistently outperforms state-of-the-art ViT methods at all pruning levels, increasing ImageNet classification accuracy by 1.3% on DeiT-T and 2.0% on DeiT-S, while reducing computational overhead by 50%. It achieves 40% speedup with improved accuracy across various ViTs.

Novendra Setyawan, Ghufron Wahyu Kurniawan, Chi-Chia Sun et al.

The perception system is a a critical role of an autonomous driving system for ensuring safety. The driving scene perception system fundamentally represents an object detection task that requires achieving a balance between accuracy and processing speed. Many contemporary methods focus on improving detection accuracy but often overlook the importance of real-time detection capabilities when computational resources are limited. Thus, it is vital to investigate efficient object detection strategies for driving scenes. This paper introduces Fast-COS, a novel single-stage object detection framework crafted specifically for driving scene applications. The research initiates with an analysis of the backbone, considering both macro and micro architectural designs, yielding the Reparameterized Attention Vision Transformer (RAViT). RAViT utilizes Reparameterized Multi-Scale Depth-Wise Convolution (RepMSDW) and Reparameterized Self-Attention (RepSA) to enhance computational efficiency and feature extraction. In extensive tests across GPU, edge, and mobile platforms, RAViT achieves 81.4% Top-1 accuracy on the ImageNet-1K dataset, demonstrating significant throughput improvements over comparable backbone models such as ResNet, FastViT, RepViT, and EfficientFormer. Additionally, integrating RepMSDW into a feature pyramid network forms RepFPN, enabling fast and multi-scale feature fusion. Fast-COS enhances object detection in driving scenes, attaining an AP50 score of 57.2% on the BDD100K dataset and 80.0% on the TJU-DHD Traffic dataset. It surpasses leading models in efficiency, delivering up to 75.9% faster GPU inference and 1.38 higher throughput on edge devices compared to FCOS, YOLOF, and RetinaNet. These findings establish Fast-COS as a highly scalable and reliable solution suitable for real-time applications, especially in resource-limited environments like autonomous driving systems

Novendra Setyawan, Ghufron Wahyu Kurniawan, Chi-Chia Sun et al.

Convolutional Neural Networks (CNNs) and Transformers have achieved remarkable success in computer vision tasks. However, their deep architectures often lead to high computational redundancy, making them less suitable for resource-constrained environments, such as edge devices. This paper introduces ParFormer, a novel vision transformer that addresses this challenge by incorporating a Parallel Mixer and a Sparse Channel Attention Patch Embedding (SCAPE). By combining convolutional and attention mechanisms, ParFormer improves feature extraction. This makes spatial feature extraction more efficient and cuts down on unnecessary computation. The SCAPE module further reduces computational redundancy while preserving essential feature information during down-sampling. Experimental results on the ImageNet-1K dataset show that ParFormer-T achieves 78.9\% Top-1 accuracy with a high throughput on a GPU that outperforms other small models with 2.56$\times$ higher throughput than MobileViT-S, 0.24\% faster than FasterNet-T2, and 1.79$\times$ higher than EdgeNeXt-S. For edge device deployment, ParFormer-T excels with a throughput of 278.1 images/sec, which is 1.38 $\times$ higher than EdgeNeXt-S and 2.36$\times$ higher than MobileViT-S, making it highly suitable for real-time applications in resource-constrained settings. The larger variant, ParFormer-L, reaches 83.5\% Top-1 accuracy, offering a balanced trade-off between accuracy and efficiency, surpassing many state-of-the-art models. In COCO object detection, ParFormer-M achieves 40.7 AP for object detection and 37.6 AP for instance segmentation, surpassing models like ResNet-50, PVT-S and PoolFormer-S24 with significantly higher efficiency. These results validate ParFormer as a highly efficient and scalable model for both high-performance and resource-constrained scenarios, making it an ideal solution for edge-based AI applications.

Ying-Hsuan Wu, Jun-Wei Hsieh, Li Xin et al.

Real-world datasets commonly exhibit noisy labels and class imbalance, such as long-tailed distributions. While previous research addresses this issue by differentiating noisy and clean samples, reliance on information from predictions based on noisy long-tailed data introduces potential errors. To overcome the limitations of prior works, we introduce an effective two-stage approach by combining soft-label refurbishing with multi-expert ensemble learning. In the first stage of robust soft label refurbishing, we acquire unbiased features through contrastive learning, making preliminary predictions using a classifier trained with a carefully designed BAlanced Noise-tolerant Cross-entropy (BANC) loss. In the second stage, our label refurbishment method is applied to obtain soft labels for multi-expert ensemble learning, providing a principled solution to the long-tail noisy label problem. Experiments conducted across multiple benchmarks validate the superiority of our approach, Label Refurbishment considering Label Rarity (LR^2), achieving remarkable accuracies of 94.19% and 77.05% on simulated noisy CIFAR-10 and CIFAR-100 long-tail datasets, as well as 77.74% and 81.40% on real-noise long-tail datasets, Food-101N and Animal-10N, surpassing existing state-of-the-art methods.

Yi-Kuan Hsieh, Jun-Wei Hsieh, Yu-Chee Tseng et al.

Traditional crowd counting networks suffer from information loss when feature maps are downsized through pooling layers, leading to inaccuracies in counting crowds at a distance. Existing methods often assume correct annotations during training, disregarding the impact of noisy annotations, especially in crowded scenes. Furthermore, the use of a fixed Gaussian kernel fails to account for the varying pixel distribution with respect to the camera distance. To overcome these challenges, we propose a Scale-Aware Crowd Counting Network (SACC-Net) that introduces a ``scale-aware'' architecture with error-correcting capabilities of noisy annotations. For the first time, we {\bf simultaneously} model labeling errors (mean) and scale variations (variance) by spatially-varying Gaussian distributions to produce fine-grained heat maps for crowd counting. Furthermore, the proposed adaptive Gaussian kernel variance enables the model to learn dynamically with a low-rank approximation, leading to improved convergence efficiency with comparable accuracy. The performance of SACC-Net is extensively evaluated on four public datasets: UCF-QNRF, UCF CC 50, NWPU, and ShanghaiTech A-B. Experimental results demonstrate that SACC-Net outperforms all state-of-the-art methods, validating its effectiveness in achieving superior crowd counting accuracy.

Munkhjargal Gochoo, Munkh-Erdene Otgonbold, Erkhembayar Ganbold et al.

With the advance of AI, road object detection has been a prominent topic in computer vision, mostly using perspective cameras. Fisheye lens provides omnidirectional wide coverage for using fewer cameras to monitor road intersections, however with view distortions. To our knowledge, there is no existing open dataset prepared for traffic surveillance on fisheye cameras. This paper introduces an open FishEye8K benchmark dataset for road object detection tasks, which comprises 157K bounding boxes across five classes (Pedestrian, Bike, Car, Bus, and Truck). In addition, we present benchmark results of State-of-The-Art (SoTA) models, including variations of YOLOv5, YOLOR, YOLO7, and YOLOv8. The dataset comprises 8,000 images recorded in 22 videos using 18 fisheye cameras for traffic monitoring in Hsinchu, Taiwan, at resolutions of 1080$\times$1080 and 1280$\times$1280. The data annotation and validation process were arduous and time-consuming, due to the ultra-wide panoramic and hemispherical fisheye camera images with large distortion and numerous road participants, particularly people riding scooters. To avoid bias, frames from a particular camera were assigned to either the training or test sets, maintaining a ratio of about 70:30 for both the number of images and bounding boxes in each class. Experimental results show that YOLOv8 and YOLOR outperform on input sizes 640$\times$640 and 1280$\times$1280, respectively. The dataset will be available on GitHub with PASCAL VOC, MS COCO, and YOLO annotation formats. The FishEye8K benchmark will provide significant contributions to the fisheye video analytics and smart city applications.

Yu-Ming Zhang, Jun-Wei Hsieh, Chun-Chieh Lee et al.

Various hand-designed CNN architectures have been developed, such as VGG, ResNet, DenseNet, etc., and achieve State-of-the-Art (SoTA) levels on different tasks. Neural Architecture Search (NAS) now focuses on automatically finding the best CNN architecture to handle the above tasks. However, the verification of a searched architecture is very time-consuming and makes predictor-based methods become an essential and important branch of NAS. Two commonly used techniques to build predictors are graph-convolution networks (GCN) and multilayer perceptron (MLP). In this paper, we consider the difference between GCN and MLP on adjacent operation trails and then propose the Redirected Adjacent Trails NAS (RATs-NAS) to quickly search for the desired neural network architecture. The RATs-NAS consists of two components: the Redirected Adjacent Trails GCN (RATs-GCN) and the Predictor-based Search Space Sampling (P3S) module. RATs-GCN can change trails and their strengths to search for a better neural network architecture. P3S can rapidly focus on tighter intervals of FLOPs in the search space. Based on our observations on cell-based NAS, we believe that architectures with similar FLOPs will perform similarly. Finally, the RATs-NAS consisting of RATs-GCN and P3S beats WeakNAS, Arch-Graph, and others by a significant margin on three sub-datasets of NASBench-201.

Bor-Shiun Wang, Jun-Wei Hsieh, Ming-Ching Chang et al.

Pooling is a simple but essential layer in modern deep CNN architectures for feature aggregation and extraction. Typical CNN design focuses on the conv layers and activation functions, while leaving the pooling layers with fewer options. We introduce the Learning Discrete Wavelet Pooling (LDW-Pooling) that can be applied universally to replace standard pooling operations to better extract features with improved accuracy and efficiency. Motivated from the wavelet theory, we adopt the low-pass (L) and high-pass (H) filters horizontally and vertically for pooling on a 2D feature map. Feature signals are decomposed into four (LL, LH, HL, HH) subbands to retain features better and avoid information dropping. The wavelet transform ensures features after pooling can be fully preserved and recovered. We next adopt an energy-based attention learning to fine-select crucial and representative features. LDW-Pooling is effective and efficient when compared with other state-of-the-art pooling techniques such as WaveletPooling and LiftPooling. Extensive experimental validation shows that LDW-Pooling can be applied to a wide range of standard CNN architectures and consistently outperform standard (max, mean, mixed, and stochastic) pooling operations.

Jun-Wei Hsieh, Ming-Ching Chang, Ping-Yang Chen et al.

Differentiable Architecture Search (DARTS) is an effective continuous relaxation-based network architecture search (NAS) method with low search cost. It has attracted significant attentions in Auto-ML research and becomes one of the most useful paradigms in NAS. Although DARTS can produce superior efficiency over traditional NAS approaches with better control of complex parameters, oftentimes it suffers from stabilization issues in producing deteriorating architectures when discretizing the continuous architecture. We observed considerable loss of validity causing dramatic decline in performance at this final discretization step of DARTS. To address this issue, we propose a Mean-Shift based DARTS (MS-DARTS) to improve stability based on sampling and perturbation. Our approach can improve bot the stability and accuracy of DARTS, by smoothing the loss landscape and sampling architecture parameters within a suitable bandwidth. We investigate the convergence of our mean-shift approach, together with the effects of bandwidth selection that affects stability and accuracy. Evaluations performed on CIFAR-10, CIFAR-100, and ImageNet show that MS-DARTS archives higher performance over other state-of-the-art NAS methods with reduced search cost.

Yu-Ming Zhang, Chun-Chieh Lee, Jun-Wei Hsieh et al.

The development of lightweight object detectors is essential due to the limited computation resources. To reduce the computation cost, how to generate redundant features plays a significant role. This paper proposes a new lightweight Convolution method Cross-Stage Lightweight (CSL) Module, to generate redundant features from cheap operations. In the intermediate expansion stage, we replaced Pointwise Convolution with Depthwise Convolution to produce candidate features. The proposed CSL-Module can reduce the computation cost significantly. Experiments conducted at MS-COCO show that the proposed CSL-Module can approximate the fitting ability of Convolution-3x3. Finally, we use the module to construct a lightweight detector CSL-YOLO, achieving better detection performance with only 43% FLOPs and 52% parameters than Tiny-YOLOv4.

Ping-Yang Chen, Jun-Wei Hsieh, Chien-Yao Wang et al.

State-of-the-art (SoTA) models have improved the accuracy of object detection with a large margin via a FP (feature pyramid). FP is a top-down aggregation to collect semantically strong features to improve scale invariance in both two-stage and one-stage detectors. However, this top-down pathway cannot preserve accurate object positions due to the shift-effect of pooling. Thus, the advantage of FP to improve detection accuracy will disappear when more layers are used. The original FP lacks a bottom-up pathway to offset the lost information from lower-layer feature maps. It performs well in large-sized object detection but poor in small-sized object detection. A new structure "residual feature pyramid" is proposed in this paper. It is bidirectional to fuse both deep and shallow features towards more effective and robust detection for both small-sized and large-sized objects. Due to the "residual" nature, it can be easily trained and integrated to different backbones (even deeper or lighter) than other bi-directional methods. One important property of this residual FP is: accuracy improvement is still found even if more layers are adopted. Extensive experiments on VOC and MS COCO datasets showed the proposed method achieved the SoTA results for highly-accurate and efficient object detection..