In this paper we address the computational inefficiency of parameter studies for automotive refrigerant circuits, which currently rely on high-fidelity but time-consuming CAE simulations. To improve time efficiency, we propose two novel machine learning models—DeepOTransformer and DeepOLSTM—trained on control values, intermediate physical parameters, power consumption, and heat exchange data. These models predict the refrigerant circuit’s physical quantities, power usage, and heat transfer rates more rapidly than conventional physics-based simulations. Compared to a baseline model, our approach achieves higher accuracy and significantly faster computation, making it a promising method for analyzing various operating conditions in automotive refrigerant systems efficiently.

Abstract Despite the significant potential of 3D object editing to impact various industries, recent research in 3D generation and editing has primarily focused on converting text and images into 3D models, often paying limited attention to the need for fine-grained control over existing 3D objects. This paper introduces a framework that uses a pre-trained regressor to enable continuous and attribute-specific modifications of both the stylistic and geometric attributes of 3D vehicle models. Here, “fine-grained control” refers to the ability to adjust specific geometric or stylistic attributes (such as roof length or perceived luxury) in a continuous and independent manner. Our method aims to preserve the identity of vehicle 3D objects and support multi-attribute editing, allowing for extensive customization while maintaining the model’s structural integrity. The framework leverages DeepSDF to obtain latent representations suitable for continuous attribute editing. Experimental results demonstrate the effectiveness of our approach in achieving detailed, controlled edits on a variety of vehicle 3D models. The code is released at https://github.com/JiangDong-miao/Vehicle_LatentEdit .

This paper presents a classification method for a Computer-Aided Diagnosis (CAD) system in a colorectal magnified Narrow Band Imaging (NBI) endoscopy. For the classification of a histologic findings, we consider an output result of a lesion endoscopic image from a pre-learned Convolutional Neural Network (CNN) as a feature vector and construct a set of Support Vector Machines (SVMs) by learning a set of the CNN feature vectors. In the video images, each frame has appearance changes such as blur, color shift, reflection of light and so on and it affects classification results. To improve the robustness of CAD system, we construct the SVM learned by multiple image sizes data sets so as to adapt to the noise peculiar to the video image. We confirmed that the proposed method achieves higher robustness, stable, and high classification accuracy in the endoscopic video image. The proposed method also can cope with differences in resolution by old and new endoscopes and perform stably with respect to the input endoscopic video image. We evaluated the proposed method on a customizable embedded DSP core implemented into a FPGA based prototyping system.

This paper presents a hardware implementation of a computer-aided diagnosis system that can classify lesions in real time for endoscopic video image. Unlike still images, video images contain many color shifts, blurs, reflections, and so on, a classification method that takes these noises into account is indispensable. In this paper, in the framework of Bag-of-Feature (BoF), we use a pre-learned convolutional Neural Network (CNN) with general images as feature extraction and Support Vector Machine (SVM) as classification. A robust classification method for endoscopic video is proposed by relearning SVM with several sets of endoscopic images. In addition, since CNN and SVM need to perform many multiplication and accumulation (MAC) operations, we implement the hardware system on a customizable DSP, which can be realize at high speed MAC operations and parallel processing with Very Long Instruction Word (VLIW). The prototyped system can be achieved real-time processing on video frame rate (30 fps @ 200 MHz). The proposed system enables not only to classify the presence or absence of cancer in the lesion but also to identify the progress of cancer and shows the clinical doctor as an objective indicator of the lesion in real time during the endoscopy examination.

In this paper, the computer-aided diagnosis system for colorectal endoscopic images is proposed. The proposed system is consisted of Convolutional Neural Network (CNN) as the feature extraction processing and Support Vector Machine (SVM) as identification processing. The proposed system is also implemented on customizable Digital Signal Processing (DSP) core: Vision P6 DSP and is demonstrated the effectiveness of a real-time recognition system by a FPGA based prototyping system, Protium™ S1.

In this paper we propose a new method for scene representation and recognition based on the concept of Region Subspaces. Each image is pre-segmented into semantically meaningful regions and local features are extracted at different scales from each such region. The Region Subspaces are the low-dimensional linear subspaces calculated from the set of local features inside each region. We also define Grassmann Matching Kernels (GMK), which extend Grassmann Kernels to be able to match simultaneously multiple subspaces. We call the resulting method, which represents image scenes through a set of Region Subspaces and matches them across different categories through the Grassmann Matching Kernel, Subspaces-of-Regions (SoR) and illustrate it on the 15-Scene dataset.

In virtual reality (VR) applications, it is often required to display a surrounding scenery with a high dynamic range (HDR) of luminance to a binocular display device such as an Oculus Rift. A surrounding scenery of an outdoor environment usually includes both bright regions, such as a sunny place and a clear sky, and dark regions, such as a shadow area. The dynamic range of the luminances between bright and dark regions are quite large.

This paper describes a method for calibrating non-overlapping cameras in a simple way: using markers on the cameras. By adding an AR (Augmented Reality) marker to a camera, we can find the transformation between the fixed AR marker and the camera's center. With such information, relative pose of cameras can be easily found as long as the marker located on them is visible. Our method consists of the following two steps: (1) use an extra camera and a chessboard to find the transformation between the AR marker and the camera center. (2) Use the information from (1) and transformation between different markers to calibrate non-overlapping cameras. Compare to other non-overlapping calibration methods, our method can work with as less as one image and does not suffer from the degenerate cases.

In this paper we propose a new local learning based regression method which utilizes ensemble-learning as a form of regularization to reduce the variance of local estimators. This makes it possible to use local learning methods even with very high-dimensional datasets. The efficacy of the proposed method is illustrated on two publicly available high-dimensional sets in comparison with several global learning methods, and it is shown that the proposed ensemble-based local learning method significantly outperforms the global ones.

Although the design of low-level local spatiotemporal features has recently led to significant improvement of performance in many action recognition applications, much less attention has been given to the equally important problem how to organize such low-level features extracted from the videos into a higher-level representation suitable to represent and discriminate between many different action classes. In this paper we propose an alternative approach to the widely-used Bag-of-Features (BoF), where instead of histograms of visual words, action sequences are represented as sets of low-dimensional linear subspaces. This results in richer and more discriminative models. In order to be able to calculate similarities between sets of subspaces we define a novel Subspace Matching Kernel. Experimental results are shown on the widely used KTH action dataset, which demonstrate the effectiveness of the proposed framework.

BACKGROUND AND OBJECTIVE: A computer-aided system for colorectal endoscopy could provide endoscopists with important helpful diagnostic support during examinations. A straightforward means of providing an objective diagnosis in real time might be for using classifiers to identify individual parts of every endoscopic video frame, but the results could be highly unstable due to out-of-focus frames. To address this problem, we propose a defocus-aware Dirichlet particle filter (D-DPF) that combines a particle filter with a Dirichlet distribution and defocus information. METHODS: We develop a particle filter with a Dirichlet distribution that represents the state transition and likelihood of each video frame. We also incorporate additional defocus information by using isolated pixel ratios to sample from a Rayleigh distribution. RESULTS: We tested the performance of the proposed method using synthetic and real endoscopic videos with a frame-wise classifier trained on 1671 images of colorectal endoscopy. Two synthetic videos comprising 600 frames were used for comparisons with a Kalman filter and D-DPF without defocus information, and D-DPF was shown to be more robust against the instability of frame-wise classification results. Computation time was approximately 88ms/frame, which is sufficient for real-time applications. We applied our method to 33 endoscopic videos and showed that the proposed method can effectively smoothen highly unstable probability curves under actual defocus of the endoscopic videos. CONCLUSION: The proposed D-DPF is a useful tool for smoothing unstable results of frame-wise classification of endoscopic videos to support real-time diagnosis during endoscopic examinations.

In this study, we have proposed an improvement for feature extraction in computer-aided diagnosis (CAD) system for colorectal endoscopic images with narrow-band imaging (NBI) magnification. Dense Scale-Invariant Feature Transform (D-SIFT) is used in the feature extraction. It is necessary to consider a trade-off between the precision of the feature extraction and speedup by the FPGA implementation for processing of real time full high definition image. In this paper, we reduced the number of dimensions for feature representation in hardware implementation purpose.

This paper proposes a method for smoothing the posterior probabilities obtained from classification results of time series input. We deal with this problem as a filtering problem with Dirichlet distribution and develop a particle filtering for this task. As a practical example of smoothing, we apply the proposed method to stabilizing NBI endoscopy recognition results over time. Experimental results demonstrate that our approach can effectively smooth highly unstable probability curves.

In this paper we propose a method for matching the scales of 3D point clouds. 3D point sets of the same scene obtained by 3D reconstruction techniques usually differ in scale. To match scales, we estimate the ratio of scales of two given 3D point clouds. By performing PCA of spin images over different scales of two point clouds, two sets of cumulative contribution rate curves are generated. Such sets of curves can be considered to characterize the scale of the given 3D point clouds. To find the scale ratio of two point clouds, we register the two sets of curves by using a variant of ICP that estimates the ratio of scales. Simulations with the Stanford bunny and experimental results with 3D reconstructions of artificial and real scenes demonstrate that the ratio of any 3D point clouds can be effectively used for scale matching.

In this paper we propose an alternative approach to the widely-used Bag-of-Features (BoF) for representing and automatically recognizing behaviors or actions in video sequences from sets of local spatio-temporal features extracted from the videos. Instead of histograms of visual words, in the proposed framework the sets of local spatio-temporal features extracted from each video are represented as low-dimensional linear subspaces, which are further othogonalized across classes to enhance their discriminability. Similarity between videos is represented in terms of Grassmann kernels defined on the subspaces of spatio-temporal features. Experimental results on a publicly available video dataset related to classifying rodent behavior demonstrate the effectiveness of the proposed framework.

In this paper we consider the case when huge datasets need to be labeled efficiently for learning. It is assumed that the data can be naturally organized into many small groups, called chunklets, each one of which contains data from the same class, and many chunklets are available from each class. Each chunklet exhibits some of the typical variation representative for the class. We investigate how active learning methods based on uncertainty sampling perform in this setting, and whether any gains can be expected in comparison with random sampling. We also propose a novel strategy for selecting which chunklets to be selected for labeling. Experiments with 7containing variation in pose, expression and illumination conditions illustrate the proposed method.

We propose an alternative approach to the widely-used Bag-of-Features (BoF) for representing objects in terms of a collection of local features extracted from their images. In this new framework, called Subspaces-of-Features (SoF), first the sets of local features extracted from the images of the objects are represented as low-dimensional linear subspaces. Then the subspaces corresponding to different categories are orthogonalized, and the similarity between subspaces corresponding to different categories is calculated using the Grassmannian distances defined through the principal angles between the subspaces. The performance of SoF is illustrated on a standard generic object recognition benchmark.

In this paper, we propose a method to select a discriminative set of image processing operations for Linear Discriminant Analysis (LDA) as an application of the use of generating matrices representing image processing operators acting on images. First we show that generating matrices can be used for formulating LDA with increasing training samples, then analyze them as image processing operators acting on 2D continuous functions for compressing many large generating matrices by using PCA and Hermite decomposition. Then we propose Linear Discriminative Image Processing Operator Analysis, an iterative method for estimating LDA feature space along with a discriminative set of generating matrices. In experiments, we demonstrate that discriminative generating matrices outperform a non-discriminative set on the ORL and FERET datasets.

In this paper, we propose a method for estimating the rotational velocity of a table tennis ball with Inverse Compositional Image Alignment, ICIA. Assuming orthogonal projection of the camera, we derive an update rule for the motion parameters. Because of the precomputation of the Hessian matrix in ICIA and the simplifying assumptions, the motion parameters estimation at each frame is very fast. We evaluate results for measuring the spins in a synthesized sequence, as well as in real image sequences of table tennis rallies.

In this paper we propose a new solution to the problem of model-based pose estimation and pose tracking, using a sequence of stereo images in which a small number of landmarks are detected on the object of interest. The landmarks are reconstructed in 3D and then the simultaneous pose and correspondence problem is solved with a novel algorithm, PC-SCO, based on stochastic combinatorial optimization. We also use a new algorithm, quaternion-based characteristic polynomial (QCP), which significantly speeds up the quaternion method used in pose alignment, making possible the tracking to be done in real-time.

In this paper we show how Grassmann distances and Grassmann kernels can be efficiently used to learn and classify face sequence videos. We propose two new methods, the Grassmann Distance Mutual Subspace Method (GD-MSM) which uses Grassmann distances to define the similarity between subspaces of images, and the Grassmann Kernel Support Vector Machine (GK-SVM), which applies two Grassmann kernels - the projection kernel and the Binet-Cauchy kernel - in a convex optimization scheme, using the Support Vector Machine (SVM) framework. GD-MSM and GK-SVM are compared in a face recognition task with several related methods using a large database of face image sequences from 100 subjects, containing expression changes related to a natural conversation setting. Additionally, we study the effect of combining all available training image sequences into a single subspace per category, in comparison with using multiple smaller subspaces, i.e. representing each category by several different subspaces, where each subspace is formed from image sequences taken under different conditions.

In this paper we propose a new method for appearance-based pose estimation, called Local Procrustes Regression (LPR). In LPR, rather than learning a map between all available training samples and pose space, as is common for appearance-based pose estimation algorithms, the pose of an unknown sample is recovered locally from a small subset of the training samples, by utilizing their inter-point distances. This is accomplished by using Procrustes analysis to align the low-dimensional image subspace generated for the neighborhood of the test sample to its corresponding part of pose space. Experimental results obtained on the Object Pose Estimation Database (OPED) indicate that the proposed method performs on par with state-of-the-art methods like support vector regression.

In this paper we propose a very efficient and low-cost method for real-time 3D object pose estimation, suitable for operation under severe changes in the illumination conditions. The method consists of a geometry-based and learning-based stage: first a geometry-based pose estimation algorithm is used to obtain pose parameter values, which are utilized as supervised information during a following learning-based stage. The efficacy of the proposed approach is illustrated on a real-world application in a head pose based facial-gesture interface for automatic wheelchair navigation.

In this paper we propose a new framework for view-invariant 3D object recognition, based on what we call Visibility Maps. A Visibility Map (VM) encodes a compact model of an arbitrary 3D object for which a set of images taken from different views is available. Representative local invariant features extracted from each image are selectively combined to form a visibility basis, in terms of which an arbitrary view of the modeled object can be represented. A metric which incorporates geometric information is also provided for comparing test images to the model, and can be used for recognition.

In this paper we propose a method for matching the scales of 3D point clouds. 3D point sets of the same scene obtained by 3D reconstruction techniques usually differ in scales. To match scales, we propose a keyscale that characterizes the scale of a given 3D point cloud. By performing PCA of spin images over different scales, a keyscale is defined as the scale that gives the minimum of cumulative contribution rate of PCA at a specific dimension of eigenspace. Simulations with the Stanford bunny and experimental results with 3D reconstructions of a real scene demonstrate that keyscales of any 3D point clouds can be uniquely found and effectively used for scale matching.

In this paper, we propose a volume rendering method using grid computing for large-scale volume data. Grid computing is attractive because medical institutions and research facilities often have a large number of idle computers. A large-scale volume data is divided into sub-volumes and the sub-volumes are rendered using grid computing. When using grid computing, different computers rarely have the same processor speeds. Thus the return order of results rarely matches the sending order. However order is vital when combining results to create a final image. Job- Scheduling is important in grid computing for volume rendering, so we use an obstacle-flag which changes priorities dynamically to manage sub-volume results. Obstacle-Flags manage visibility of each sub-volume when line of sight from the view point is obscured by other sub-volumes. The proposed dynamic job-scheduling based on visibility substantially increases efficiency. Our dynamic job-scheduling method was implemented on our university's campus grid and we conducted comparative experiments, which showed that the proposed method provides significant improvements in efficiency for large-scale volume rendering.

We have developed a head gesture controlled electric wheelchair system to aid persons with severe disabilities. Realtime range information obtained from a stereo camera is used to locate and segment the face images of the user from the sensed video. We use an Isomap based nonlinear manifold learning map of facial textures for head pose estimation. Our system is a non-contact vision system, making it much more convenient to use. The user is only required to gesture his/her head to command the wheelchair. To overcome problems with a non responding system, it is necessary to notify the user of the exact system state while the system is in use. In this paper, we explore the use of vibrotactile rendering of head gestures as feedback. Three different feedback systems are developed and tested, audio stimuli, vibrotactile stimuli and audio plus vibrotactile stimuli. We have performed user tests to study the usability of these three display methods. The usability studies show that the method using both audio plus vibrotactile response outperforms the other methods (i.e. audio stimuli, vibrotactile stimuli response).

We propose a new method for unsupervised face recognition - VQ-faces, which operates on a sequential stream of face images and is able to handle both frontal and side-view faces at the same time. The method consists of two parts: in the first part, the VQ-faces are calculated as prototype vectors of local areas in image-space, coding for different face-views (i.e. a "view codebook" is generated), while in the second part the fact that each face-sequence corresponds to a single person (temporal constraint) is used to cluster the different sequences into face categories, using a combinatorial optimization process which maximizes an objective function, reducing the global representation error (distortion) by the VQ-faces as sequences are grouped together. The method was tested on real-world data gathered over a period of several months and including both frontal and side-view faces from 17 different subjects, achieving correct self-organization rate of 85.4%. The applicability of the proposed method is not limited to face recognition - it can be easily applied to other problems involving multi-view object recognition.

We propose a new method for unsupervised face recognition from time-varying sequences of face images obtained in real-world environments. Two types of forces, attraction and repulsion, operate across the spatio-temporal facial manifolds, to autonomously organize the data without relying on any category-specific information provided in advance. Experiments with real-world data gathered over a period of several months and including both frontal and side-view faces were used to evaluate the method and encouraging results were obtained The proposed method can be used in video surveillance systems or for content-based information retrieval.

We propose a novel method for unsupervised face recognition from time-varying sequences of face images obtained in real-world environments. The method utilizes the higher level of sensory variation contained in the input image sequences to autonomously organize the data in an incrementally built graph structure, without relying on category-specific information provided in advance. This is achieved by "chaining" together similar views across the spatio-temporal facial manifolds by two types of connecting edges depending on a local measure of similarity. Experiments with real-world data gathered over a period of several months and including both frontal and side-view faces from 17 different subjects were used to test the method, achieving a correct self-organization rate of 88.6%. The proposed method can be used in video surveillance systems or for content-based information retrieval.