• Journal of Information Processing Systems
• /
• v.18 no.6
• /
• pp.794-802
• /
• 2022

The recovery of reasonable depth information from different scenes is a popular topic in the field of computer vision. For generating depth maps with better details, we present an efficacious monocular depth prediction framework with coordinate attention and feature fusion. Specifically, the proposed framework contains attention, multi-scale and feature fusion modules. The attention module improves features based on coordinate attention to enhance the predicted effect, whereas the multi-scale module integrates useful low- and high-level contextual features with higher resolution. Moreover, we developed a feature fusion module to combine the heterogeneous features to generate high-quality depth outputs. We also designed a hybrid loss function that measures prediction errors from the perspective of depth and scale-invariant gradients, which contribute to preserving rich details. We conducted the experiments on public RGBD datasets, and the evaluation results show that the proposed scheme can considerably enhance the accuracy of depth prediction, achieving 0.051 for log10 and 0.992 for δ<1.25³ on the NYUv2 dataset.

• 제어로봇시스템학회:학술대회논문집
• /
• 2005.06a
• /
• pp.383-388
• /
• 2005

Depth recovery in robot vision is an essential problem to infer the three dimensional geometry of scenes from a sequence of the two dimensional images. In the past, many studies have been proposed for the depth estimation such as stereopsis, motion parallax and blurring phenomena. Among cues for depth estimation, depth from lens translation is based on shape from motion by using feature points. This approach is derived from the correspondence of feature points detected in images and performs the depth estimation that uses information on the motion of feature points. The approaches using motion vectors suffer from the occlusion or missing part problem, and the image blur is ignored in the feature point detection. This paper presents a novel approach to the defocus technique based depth from lens translation using sequential SVD factorization. Solving such the problems requires modeling of mutual relationship between the light and optics until reaching the image plane. For this mutuality, we first discuss the optical properties of a camera system, because the image blur varies according to camera parameter settings. The camera system accounts for the camera model integrating a thin lens based camera model to explain the light and optical properties and a perspective projection camera model to explain the depth from lens translation. Then, depth from lens translation is proposed to use the feature points detected in edges of the image blur. The feature points contain the depth information derived from an amount of blur of width. The shape and motion can be estimated from the motion of feature points. This method uses the sequential SVD factorization to represent the orthogonal matrices that are singular value decomposition. Some experiments have been performed with a sequence of real and synthetic images comparing the presented method with the depth from lens translation. Experimental results have demonstrated the validity and shown the applicability of the proposed method to the depth estimation.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.12 no.6
• /
• pp.395-403
• /
• 2017

Image processing for object detection and identification has been studied for supply chain management application with various approaches. Among them, feature pointed detection algorithm is used to track an object or to recognize a position in automated supply chain systems and a depth image based feature point detection is recently highlighted in the application. The result of feature point detection is easily influenced by image noise. Also, the depth image has noise itself and it also affects to the accuracy of the detection results. In order to solve these problems, we propose a novel hybrid filtering mechanism for depth image based feature point detection, it shows better performance compared with conventional hybrid filtering mechanism.

• Journal of KIISE:Software and Applications
• /
• v.31 no.5
• /
• pp.617-624
• /
• 2004

This paper describes a new facial feature localization method that uses Adjacent Depth Differences(ADD) in 3D facial surface. In general, human recognize the extent of deepness or shallowness of region relatively, in depth, by comparing the neighboring depth information among regions of an object. The larger the depth difference between regions shows, the easier one can recognize each region. Using this principal, facial feature extraction will be easier, more reliable and speedy. 3D range images are used as input images. And ADD are obtained by differencing two range values, which are separated at a distance coordinate, both in horizontal and vertical directions. ADD and input image are analyzed to extract facial features, then localized a nose region, which is the most prominent feature in 3D facial surface, effectively and accurately.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.8
• /
• pp.2068-2082
• /
• 2023

With the rapid development of deep learning, Depth Map Super-Resolution (DMSR) method has achieved more advanced performances. However, when the upsampling rate is very large, it is difficult to capture the structural consistency between color features and depth features by these DMSR methods. Therefore, we propose a color-image guided DMSR method based on iterative depth feature enhancement. Considering the feature difference between high-quality color features and low-quality depth features, we propose to decompose the depth features into High-Frequency (HF) and Low-Frequency (LF) components. Due to structural homogeneity of depth HF components and HF color features, only HF color features are used to enhance the depth HF features without using the LF color features. Before the HF and LF depth feature decomposition, the LF component of the previous depth decomposition and the updated HF component are combined together. After decomposing and reorganizing recursively-updated features, we combine all the depth LF features with the final updated depth HF features to obtain the enhanced-depth features. Next, the enhanced-depth features are input into the multistage depth map fusion reconstruction block, in which the cross enhancement module is introduced into the reconstruction block to fully mine the spatial correlation of depth map by interleaving various features between different convolution groups. Experimental results can show that the two objective assessments of root mean square error and mean absolute deviation of the proposed method are superior to those of many latest DMSR methods.

• Journal of KIISE:Software and Applications
• /
• v.29 no.11
• /
• pp.818-825
• /
• 2002

Depth information is one of the most important factor for the recognition of a digital face image. Range images are very useful, when comparing one face with other faces, because of implicating depth information. As the processing for the whole fare produces a lot of calculations and data, face images ran be represented in terms of a vector of feature descriptors for a local area. In this paper, depth areas of a 3 dimensional(3D) face image were extracted by the contour line from some depth value. These were resampled and stored in consecutive location in feature vector using multiple feature method. A comparison between two faces was made based on their distance in the feature space, using Euclidian distance. This paper reduced the number of index data in the database and used fewer feature vectors than other methods. Proposed algorithm can be highly recognized for using local depth information and less feature vectors or the face.

• The Journal of the Korea Contents Association
• /
• v.18 no.11
• /
• pp.447-454
• /
• 2018

This paper proposes a depth image generation algorithm of stereo images using a deep learning model composed of a CNN (convolutional neural network). The proposed algorithm consists of a feature extraction unit which extracts the main features of each parallax image and a depth learning unit which learns the parallax information using extracted features. First, the feature extraction unit extracts a feature map for each parallax image through the Xception module and the ASPP(Atrous spatial pyramid pooling) module, which are composed of 2D CNN layers. Then, the feature map for each parallax is accumulated in 3D form according to the time difference and the depth image is estimated after passing through the depth learning unit for learning the depth estimation weight through 3D CNN. The proposed algorithm estimates the depth of object region more accurately than other algorithms.

• The KIPS Transactions:PartB
• /
• v.17B no.1
• /
• pp.55-62
• /
• 2010

Human visual attention system has a remarkable ability to interpret complex scenes with the ease and simplicity by selecting or focusing on a small region of visual field without scanning the whole images. In this paper, a novel feature map generation and integration method for attention based visual information processing system is proposed. The depth information obtained from a stereo pair of images is exploited as one of spatial visual features to form a set of topographic feature maps in our approach. Comparative experiments show that correct detection rate of visual attention regions improves by utilizing depth feature compared to the case of not using depth feature.

• Advances in robotics research
• /
• v.2 no.1
• /
• pp.33-44
• /
• 2018

This paper presents a novel feature detection algorithm called depth-hybrid speeded-up robust features (DH-SURF) augmented by depth information in the speeded-up robust features (SURF) algorithm. In the keypoint detection part of classical SURF, the standard deviation of the Gaussian kernel is varied for its scale-invariance property, resulting in increased computational complexity. We propose a keypoint detection method with less variation of the standard deviation by using depth data from a red-green-blue depth (RGB-D) sensor. Our approach maintains a scale-invariance property while reducing computation time. An RGB-D simultaneous localization and mapping (SLAM) system uses a feature extraction method and depth data concurrently; thus, the system is well-suited for showing the performance of the DH-SURF method. DH-SURF was implemented on a central processing unit (CPU) and a graphics processing unit (GPU), respectively, and was validated through the real-time RGB-D SLAM.

• IEIE Transactions on Smart Processing and Computing
• /
• v.4 no.6
• /
• pp.379-383
• /
• 2015

This paper presents an approach to detect object distance with the use of the recently developed low-cost Kinect sensor. The technique is based on Kinect color depth-image processing and can be used to design various computer-vision applications, such as object recognition, video surveillance, and autonomous path finding. The proposed technique uses keypoint feature detection in the Kinect depth image and advantages of depth pixels to directly obtain the feature distance in the depth images. This highly reduces the computational overhead and obtains the pixel distance in the Kinect captured images.