ABSTRACT. 在電腦 3D 動畫中，要使模型能夠活動，必須要為模型建立骨架。將角色模型與骨架結合後匯入動作檔，動作檔定義骨架關節的運動，藉此帶動角色蒙皮的變形，此流程稱為骨骼動畫（Rigging/Skeletal Animation）。骨骼動畫以往的作法是由角色建模師以人工方式建立骨架、並人工調整蒙皮錯誤的變形。所以針對大量模型進行骨骼動畫時，耗費大量人力與時間。因此，自動骨骼動畫技術與機器學習骨骼動畫技術開始發展。上述技術目前會骨架錯置的狀況，而且無法產生細部骨架（如手指關節）進行骨骼動畫。本研究計畫目標在提高 RigNet 骨骼生成的細節，使其能生成細部手指骨架。

ABSTRACT. 本研究提出了基於深度學習與標記分水嶺變換 (Marker-based Watershed Transform) 之方法, 以解決從錐狀射束電腦斷層掃描 (Cone Beam Computed Tomography, CBCT) 影像進行準確地牙齒實例分割之問題。由於 CBCT 影像具有牙齒結構複雜、牙齒周圍邊界不清晰以及金屬偽影等因素，手動進行牙齒實例分割是一項耗時且具有挑戰性的任務。為了解決這個問題，首先，本研究提出了名為 CSWin UNETR 之三維深度學習模型，此為 3D CSWin Transformer 與 3D UNETR 組成之編碼器解碼器網路。並將 CSWin UNETR 組成雙分支網路以同時精確地分割牙齒區域與中心。 最後，透過標記分水嶺變換實現牙齒實例分割。實驗結果表明，提出之方法能於金屬植體、智齒與缺牙的影像中有效地分割單個牙齒。此外，所提出之 CSWin UNETR 在牙齒區域分割任務中，取得之 Dice 相似係數為 0.9640，95% Hausdorff 距離為 1.0414 mm，其定量分析結果皆優於現有之優秀架構如 Swin UNETR、UNETR 與 U-Net 等。結果表明，提出之方法能對 CBCT 影像進行準確地牙齒實例分割，以達到建構三維牙齒模型之目的。

ABSTRACT. 文字生成圖像的課題上，除了生成對抗網路手法之外，近年來擴散模型(Diffusion Model)的大幅進步，達成了一些令人印象深刻的研究成果並對生成圖像的發展有重大貢獻。本文使用了預訓練的穩定擴散模型(Stable Diffusion Model)並對模型進行微調，讓用戶輸入理想的服裝文本並生成具觀賞性及創造性的服裝圖像。此模型對於文本的輸入具有相當好的敏銳性，能對文本中對服裝的細節以及輪廓描述，產生相對應的成果。 衣著品味常常是我們給人的第一印象，為了給予他人一個深刻且良好的印象，我們關心著自己的穿著和外表，於是人們不斷尋找心中所想的適合自己的完美服裝。因此我們希望透過模型文本生成圖像的能力將想法可視化，在現實中以圖像的形式看到實際的服裝樣貌，使每個人都能成為自己的服裝設計師，更容易地找尋到自己理想的服裝。 於本文的實驗中，使用自行蒐集的流行服飾資料集，其內容為真實的服裝，由高清質量的圖像組成，文本則是透過專業設計師的描述而來，主要提供了關於輪廓以及細節的描述。 此外，本文為了加強使用者所輸入的文本質量，製作了一個基於Ｗord2vec的推薦服裝描述模型。這個模型可以自動從用戶所輸入的句子去推薦出更多關於服裝的細節以及相對應的用詞，幫助使用者更加精確地描述他們所需要的服裝風格。透過應用這個模型，可以提高服裝生成的品質，並且讓使用者更加便利地輸入他們所需要的服裝文本。

ABSTRACT. 樂高公司持續推出新的系列及不同種類的磚，其多樣性深受大人小孩的喜愛，相對較少有研究進行探討新推出的樂高系列，但也會有與其他領域，像是拼貼、排列、像素化問題等或與樂高研究議題相關。像素化藝術方面的研究一直以來都非常受歡迎，但主要限制是基本單位都為方形，無法保留輸入圖形中圓滑的特徵。本研究的樂高豆豆系列則是將畫素以實體元件表現出來，而除了方形的磚，也有一些帶有較圓滑邊緣的磚。但當我們要以人工的方式去拼一個形狀時，在磚形狀及顏色的選擇上，就會花費非常多的時間，如果要拼的東西越大，所花費的時間也就更久，會浪費許多勞力和時間。為了解決這些問題，本研究首先嘗試將編碼器-解碼器架構的網路與二維樂高平板磚建構相結合。以樂高平板磚中的方形磚及帶圓滑邊緣的磚轉換為圖片作為輸入，並透過設計以感知損失為基礎損失函數，將現有的照片馬賽克神經網路研究延伸至樂高平板磚的組合問題。同時，我們也針對輸入圖形及生成的圖片，做一系列的比較及分析，證明此系統的有效性。

ABSTRACT. In this contribution, we empirically investigated the effect of small talk on the users' non-verbal behaviors and emotions when users interacted with a crowd of virtual humans (VHs) with positive behavioral dispositions. Users were tasked with collecting items in a virtual marketplace via natural speech-based dialogue with a crowd of virtual pedestrians and vendors. The users were able to engage in natural speech-based conversation in a predefined corpus of small talk content that covered various commonplace small talk topics such as conversations about the weather, general concerns, and entertainment based on similar real-life situations. For instance, the VHs with the small talk ability would ask the users some simple questions to make small talk or remind the users of their belongings. We conducted a between-subjects empirical evaluation to investigate whether the user behaviors and emotions were different between a small talk condition and a non-small talk condition, and examined gender effects of the participants. We collected objective and subjective measures of the users to analyze users' emotions and social interaction behaviors, when in conversation with VHs that either possessed small-talk capability or not, besides task or goal oriented dialogue capabilities. Our result revealed that the VHs with small talk capability could alter the emotions and non-verbal behaviors of the users. Furthermore, the non-verbal behaviors between female and male participants differed greatly in the presence or absence of small talk.

ABSTRACT. This paper presents a within-subject study to investigate the effects of leading and following behaviors on user visual attention behaviors when collaborating with a virtual agent (VA) during performing transportation tasks in immersive virtual environments. The task was to carry a target object from a location to a predefined location. There were two conditions, namely leader VA (LVA) and follower VA (FVA). The leader gave instructions to the follower to perform actions. In the FVA condition, users played the leader role, while they played the follower role in the LVA condition. The users and the VA communicated via spoken language. During the experiment, participants wore a head-mounted display and performed real walking in a room. In each condition, each participant performed 20 trials of object transportation for different types of objects. Our preliminary results revealed significant differences in the user visual attention behaviors between the follower and leader VA conditions during the transportation tasks.

ABSTRACT. We present a colorization network that generates flat-color icons according to given sketches and semantic colorization styles. Our network contains a style-structure disentangled colorization module and a normalizing flow. The colorization module transforms a paired sketch image and style image into a flat-color icon. To enhance network generalization and the quality of icons, we present a pixel-wise decoder, a global style code, and a contour loss to reduce color gradients at flat regions and increase color discontinuity at boundaries. The normalizing flow maps Gaussian vectors to diverse style codes conditioned on the given semantic colorization label. This conditional sampling enables users to control attributes and obtain diverse colorization results. Compared to previous methods built upon conditional generative adversarial networks, our approach enjoys the advantages of both high image quality and diversity. To evaluate its effectiveness, we compared the flat-color icons generated by our approach and recent colorization and image-to-image translation methods on various conditions. Experiment results verify that our method out- performs state-of-the-arts qualitatively and quantitatively.

ABSTRACT. This paper investigated to what extent social interactions and empathy of users could be induced when different control mechanisms were used in an asymmetric collaboration. We conducted a user study to explore the user experience under one decentralized and two centralized control conditions via using the proposed two-player asymmetric collaborative bodily play, LAGH, which supports perspective-taking through the integration with the first- and second-perspectives and shared objects. The two players have complementary views and controls to each other in an immersive environment. The results indicate that participant pairs were encouraged by the asymmetric collaboration interface to share their emotional and physiological perspectives with each other. When their control abilities were balanced, they were more motivated to perform information sharing and interact with each other, thereby enhancing closeness and stimulating empathy. Furthermore, users could improve the collaboration efficiency.

ABSTRACT. 單細胞辨識常作為生物醫學的研究對象並透過三維共軛焦顯微影像可觀察三維空間中細胞生長、分裂、分化的狀況，因此單細胞自動辨識正確與否，影響細胞生長的使用移動、消滅和分裂等分析結果。而單細胞因彼此細胞相近，導致單細胞自動辨識不易，尤其是影像堆疊方向解析度較低之三維共軛焦顯微影像更是如此。在深度學習技術最近也常用於辨識分割細胞範圍，但仍不易辨識相近單細胞相連的問題。本研究提出多方向細胞中心偵測透過不同方向的細胞中心之偵測並整合不同方向所觀測到的細胞中心範圍得到正確，藉以改善三維共軛焦顯微影像中，三維接連單細胞的分割問題。實驗結果顯示，此多方向偵測細胞中心方式，能較精確分割三維空間中的相連單細胞，以利於後續進行細胞之相關分析。

ABSTRACT. 現在AR眼鏡的相關應用不斷地推陳出新，其中，最常被使用於日常生活中的應用，莫過於訊息接收或是文章閱讀。AR眼鏡打破了VR眼鏡的限制，讓用戶可以同時閱讀，也可以同時感知到周遭情況，而不是像VR眼鏡會遮擋用戶對於真實世界的視界。而且不同的文字呈現方式也影響著閱讀的效率和對於用戶的身體副作用，像是頭暈、眼睛痛等。所以，我們使用Moverio Bt-300這款AR眼鏡對五位參與者進行文本閱讀實驗，通過模擬六種AR眼鏡的使用情況，這六種情況可以分為三大類，每一大類共兩種不同的使用情況，第一大類:文字所呈現的地方，讓文字出現在AR眼鏡裡面或是藉由特定花紋的板子讓文字投射在板子上面，第二大類:文字呈現的多寡，讓文字一次填滿AR眼鏡的可視範圍內，或是一次只出現一行文字，第三大類:文字的操控模式，讓文字是以全自動的方式播放或是可以由操控者手動改變下一篇文章。以上的情況是為了測試在現實和虛擬世界中閱讀文本的副作用大小，以追求最佳的文字閱讀模式。詳細的解說可以從我們的研究中近一步了解AR眼鏡使用情況的設計想法。

ABSTRACT. The combination of automation equipment and intelligence has become the trend of the automation industry. Among them, AOI (Automatic optical inspection) applications are applied in the field of automation due to their advantages such as not being limited by working hours, rapid detection, and low labor costs. For the case of mass production, it can solve the shortage of operator configuration, learning cost and production efficiency. Therefore, this study uses the AOI system of YoloV4 (You Only Look Once Version4) architecture for sample detection, and uses Ethernet communication architecture to integrate six-axis robotic arms, PLC (Programmable Logic Controller), industrial computers and other equipment. It has intelligent functions such as automatic scheduling, statistical production reports, real-time monitoring, etc., and a GUI (Graphical User Interface) graphical interface is designed to reduce personnel learning costs, operation difficulty and improve equipment utilization, thereby forming a highly efficient smart machine system.

ABSTRACT. 在近年來，握力被證實是一種重要的健康指標，握力測試可用來評估個人的肌肉量，同時，握力影響著日常上肢的施力及活動度，包含抓、握和扭等等需要上肢施力的動作。握力經研究後也跟罹患糖尿病風險有一定的關係，因此握力其實跟一般人的健康程度亦息息相關。然而，以往在握力上的相關運動並未在大多數人中受到重視與實施，且目前相關研究的領域中，大多數都屬醫學領域。但是，握力是可以藉由日常的運動及鍛鍊來得到提升的。所以，在本篇研究裡我們將其結合遊戲化，遊戲化可提升的益處很多，不僅可以增加使用者在鍛鍊中的樂趣、提高使用者動力及參與度，也不受天氣、地形及環境安全影響。我們製作了鍛鍊握力的運動遊戲於Unity3D，遊戲製作完後安裝並執行於安卓系統(Android)的手機上，使用 Google Cardboard實現VR系統並結合IoT及握力器來進行遊戲體驗。同時加入其他遊戲化設計，製作出一款於不同系統間的遊戲雛型整合及開發。

ABSTRACT. In this article, we propose a texture mapping algorithm for gluing 2D patterns on voxel-based models’ surfaces. Since voxel-based models are composed of voxels, their surfaces are irregular and basic geometrical information, like normal and tangent vectors, are absent from their digital representations. We cannot rely on connectivity and geometrical information to parametrize the surface of a voxel-based model. Instead, we derive an automatic mapping procedure to parametrize these voxels by using only their 3D coordinates. As a result, texture mapping on volumetric models can be accomplished by using little human interference. The proposed method utilizes Self-Organizing Map (SOM) to compromise the topological differences between 2D textures and voxel-model-surfaces. When texture-mapping a volumetric model, we select a SOM lattice as the intermediate geometry at first. Then, the SOM lattice is undertaken an unsupervised training to fit the shape of the model. Subsequently, another unsupervised training is triggered to parameterize the SOM lattice such that each SOM node is associated with a pair of texture coordinates. These two training procedures transform the SOM into a bridge, connecting the model’s surface and the texture map. The surface voxels of the model can thus acquire their texture coordinates based on the SOM.

ABSTRACT. For over the past three decades, researchers in the field of interactive 3D computer graphics have intensively studied the issue of Level-Of-Detail (LOD) control. A great number of techniques for balancing the rendering efficiency and fidelity demands were proposed. Among these techniques, a class of techniques called the discrete or static LOD technique is commonly adopted in real-time 3D graphics applications such as virtual reality (VR) and 3D computer games for its simplicity and easiness of implementation. However, according to previous works, the discrete LOD-based representations usually suffered from the problem of loading latency, which caused the "popping effect" that resulted in a visual disturbance when the system switches LOD. To cope with this problem, we proposed a novel approach called common vertex LOD. In this paper, we proposed a Common Vertex Buffer LOD representation, the algorithms for generating such LOD meshes, and a primary investigation of the effectiveness in reducing the loading latency of the LOD meshes.

ABSTRACT. Virtual reality (VR) technology has developed rapidly in recent years. However, the design of effective VR scenarios, which is labor-intensive and time-consuming, remains a challenge. In response to this challenge, this paper presents a novel approach to developing interactive VR scenarios using mobile VR tangible user interfaces and 2D/3D layered modularized visual objects. The main objectives of the approach are that the use of modularized visual objects allows for flexible and scalable design of virtual environments. In addition, the use of mobile VR is intended to enhance user engagement and immersion by creating tailored to specific scenario requirements and to provide a sustainable, personal and safe way for users to interact with the VR environment. The proposed approach is demonstrated through the development of a design prototype and user studies are conducted to evaluate the effectiveness of the approach. The results indicate that the integration of mobile VR tangible user interfaces and the modularization of visual objects can effectively support users in the interactive design of VR scenarios. This approach has the potential to be applied to a variety of VR applications and can make a significant contribution to the field of VR research and development.

ABSTRACT. This paper presents a prototype of a virtual reality exercise game that uses machine learning to control content generation and game personalization. The game aims to provide a personalized workout experience for users by generating content that is tailored to their individual grip training level, interests and preferences. Genetic algorithms and artificial intelligence neural network algorithms are used to analyze user data such as their biometrics, workout history and feedback to generate challenging but achievable personalized workout routines. The game also incorporates gamification designs to promote engagement and motivation, such as NPC, score, rewards and so on. The prototype was evaluated through user research, which showed that participants found the content motivating and enjoyable. The results suggest that using machine learning for content generation and personalization can improve the user experience and encourage adherence to the training application in a virtual reality environment.

ABSTRACT. The shoe industry in Taiwan is facing a dilemma of industrial transformation, which requires driving product innovation, service innovation, and new technology. This paper aims to address this challenge by leveraging the growing trend of wearable devices in healthcare and the commercial application of NB-IoT. To achieve this goal, we have developed smart healthcare shoes equipped with pressure sensors in the insole. The data collected from these sensors is transmitted via NB-IoT to a cloud storage for establishing a plantar pressure database. We have also used artificial intelligence and deep learning algorithms to classify and cluster related data and build a time-series model to develop the Intelligent Analysis System of Foot Pressure (IASFP). The experimental results demonstrate that our proposed smart healthcare shoes have an accuracy and precision rate of over 90% for all plantar pressure measurements. Therefore, our approach offers a promising solution to the challenges of industrial transformation faced by the shoe industry in Taiwan.

ABSTRACT. Measurement of plant root system architecture (RSA) traits is an important task for botany. Usually, the botanists put the plants in a transparent gel container for easy observation. Under this configuration, an easy-to-use way to measure RSA traits is to take images for observation. However, in single-view-angle 2D image often has problems such as occlusion and lack of depth information, so it is not convenient for measurement. Therefore, the reconstruction of the 3D root model from multi-view-angle images is a better solution. Under the above-mentioned planting configuration, the refractive distortion problem usually arises. This will lead to serious distortion of model reconstruction, so in this paper, a method based on ray tracing for correcting refraction distortion of objects in cylindrical containers is proposed.