• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.6
• /
• pp.2044-2059
• /
• 2022

Recent works have shown that the geometric constraint can be harnessed to boost the performance of CNN-based camera localization. However, the existing strategies are limited to imposing image-level constraint between pose pairs, which is weak and coarse-gained. In this paper, we introduce a pixel-level epipolar geometry constraint to vanilla localization framework without the ground-truth 3D information. Dubbed EpiLoc, our method establishes the geometric relationship between pixels in different images by utilizing the epipolar geometry thus forcing the network to regress more accurate poses. We also propose a variant called EpiSingle to cope with non-sequential training images, which can construct the epipolar geometry constraint based on a single image in a self-supervised manner. Extensive experiments on the public indoor 7Scenes and outdoor RobotCar datasets show that the proposed pixel-level constraint is valuable, and helps our EpiLoc achieve state-of-the-art results in the end-to-end camera localization task.

• Proceedings of the KSRS Conference
• /
• 1999.11a
• /
• pp.182-187
• /
• 1999

Stereo matching is one of the most crucial parts in DEM generation. Naive stereo matching algorithms often create many holes and blunders in a DEM and therefore a carefully designed strategy must be employed to guide stereo matching algorithms to produce “good” 3D information. In this paper, we describe one such a strategy designed by the use of scene geometry, in particular, the epipolarity for generation of a DEM from linear pushbroom images. The epipolarity for perspective images is a well-known property, i.e., in a stereo image pair, a point in the reference image will map to a line in the search image uniquely defined by sensor models of the image pair. This concept has been utilized in stereo matching by applying epipolar resampling prior to matching. However, the epipolar matching for linear pushbroom images is rather complicated. It was found that the epipolarity can only be described by a Hyperbola- shaped curve and that epipolar resampling cannot be applied to linear pushbroom images. Instead, we have developed an algorithm of incorporating such epipolarity directly in least squares correlation matching. Experiments showed that this approach could improve the quality of a DEM.

• Journal of Korean Society for Geospatial Information Science
• /
• v.15 no.4
• /
• pp.81-88
• /
• 2007

The geometry of satellite image captured by linear CCD sensor is different from that of frame camera image. The fact that the exterior orientation parameters for satellite image with linear CCD sensor varies from scan line by scan line, causes the difference of image geometry between frame and linear CCD sensor. Therefore, we need the epipolar geometry for linear CCD image which differs from that of frame camera image. In this paper, we proposed a method of resampling linear CCD satellite image in epipolar geometry under the assumption that image is not formed in perspective projection but in parallel projection, and the sensor model is a 2D affine sensor model based on parallel projection. For the experiment, IKONOS stereo images, which are high resolution linear CCD images, were used and tested. As results, the spatial accuracy of 2D affine sensor model is investigated and the accuracy of epipolar resampled image with RFM was presented.

• 제어로봇시스템학회:학술대회논문집
• /
• 2003.10a
• /
• pp.1121-1125
• /
• 2003

In this paper we deal with visual servoing that can control a robot arm with a camera using information of images only, without estimating 3D position and rotation of the robot arm. Here it is assumed that the robot arm is calibrated and the camera is uncalibrated. We use a pinhole camera model as the camera one. The essential notion can be show, that is, epipolar geometry, epipole, epipolar equation, and epipolar constrain. These play an important role in designing visual servoing. For easy understanding of the proposed method we first show a design in case of the calibrated camera. The design is constructed by 4 steps and the directional motion of the robot arm is fixed only to a constant direction. This means that an estimated epipole denotes the direction, to which the robot arm translates in 3D space, on the image plane.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.36 no.3
• /
• pp.127-134
• /
• 2018

Electric transmission towers are facilities to transport electrical power from a plant to an electrical substation. The towers are connected using power lines that are installed with a proper sag by loosening the cable to lower the tension and to secure the sufficient clearance from the ground or nearby objects. The power line sag may extend over the tolerance due to the weather such as strong winds, temperature changes, and a heavy snowfall. Therefore the periodical mapping of the power lines is required but the poor accessibility to the power lines limit the work because most power lines are placed at the mountain area. In addition, the manual mapping of the power lines is also time-consuming either using the terrestrial surveying or the aerial surveying. Therefore we utilized multiple overlapping images acquired from a low-cost drone to automatically reconstruct the power lines in the object space. Two overlapping images are selected for epipolar image resampling, followed by the line extraction for the resampled images and the redundant images. The extracted lines from the epipolar images are matched together and reconstructed for the power lines primitive that are noisy because of the multiple line matches. They are filtered using the extracted line information from the redundant images for final power lines points. The experiment result showed that the proposed method successfully generated parabolic curves of power lines by interpolating the power lines points though the line extraction and reconstruction were not complete in some part due to the lack of the image contrast.

• Journal of Broadcast Engineering
• /
• v.18 no.1
• /
• pp.31-42
• /
• 2013

This paper proposes a photogrammetric rectification method based on Bayesian approach as a method that eliminates vertical parallax between stereo images to minimize visual fatigue of 3D contents. The image rectification consists of two phases; geometry estimation and epipolar transformation. For geometry estimation, coplanarity-based relative orientation algorithm was used in this paper. To ensure robustness for mismatch and localization error occurred by automation of tie point extraction, Bayesian approach was applied by introducing several prior constraints. As epipolar transformation perspective transformation was used based on condition of collinearity to minimize distortion of result images and modification for input images. Other algorithms were compared to evaluate performance. For geometry estimation, traditional relative orientation algorithm, 8-points algorithm and stereo calibration algorithm were employed. For epipolar transformation, Hartley algorithm and Bouguet algorithm were employed. The evaluation results showed that the proposed algorithm produced results with high accuracy, robustness about error sources and minimum image modification.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.35S no.12
• /
• pp.84-92
• /
• 1998

In this paper, we propose the novel cost functions for finding the disparity between the left and the right images in the stereo matching problem. The dynamic programming method is used in solving the stereo matching problem by Cox et al[10]. In the reference[10], only the intensity of the pixels in the epipolar line is used as the cost functions to find the corresponding pixels. We propose the two new cost functions. The information of the slope of the pixel is introduced to the constraints in determining the weights of intensity and direction(the historical information). The pixels with the higher slope are matched mainly by the intensity of pixels. As the slope becomes lower, the matching is performed mainly by the direction. Secondly, the disparity information of the previous epipolar line the pixel is used to find the disparity of the current epipolar line. If the pixel in the left epipolar line, $p-i$ and the pixel in the right epipolar line, $p-j$ satisfy the following conditions, the higher matching probability is given to the pixels, $p-i$ and $p-j$. i) The pixels, $p-i$ and $p-j$ are the pixles on the edges in the left and the right images, respectively. ⅱ) For the pixels $p-k$ and $p-l$ in the previous epipolar line, $p-k$and $p-l$ are matched and are the pixels on the same edge with $p-i$ and $p-j$, respectively. The proposed method compared with the original method[10] finds the better matching results for the test images.

• ETRI Journal
• /
• v.39 no.1
• /
• pp.97-107
• /
• 2017

A fast and accurate building-level visual place recognition method built on an image-retrieval scheme using street-view images is proposed. Reference images generated from street-view images usually depict multiple buildings and confusing regions, such as roads, sky, and vehicles, which degrades retrieval accuracy and causes matching ambiguity. The proposed practical database refinement method uses informative reference image and keypoint selection. For database refinement, the method uses a spatial layout of the buildings in the reference image, specifically a building-identification mask image, which is obtained from a prebuilt three-dimensional model of the site. A global-positioning-system-aware retrieval structure is incorporated in it. To evaluate the method, we constructed a dataset over an area of $0.26km^2$. It was comprised of 38,700 reference images and corresponding building-identification mask images. The proposed method removed 25% of the database images using informative reference image selection. It achieved 85.6% recall of the top five candidates in 1.25 s of full processing. The method thus achieved high accuracy at a low computational complexity.

• ETRI Journal
• /
• v.33 no.4
• /
• pp.537-546
• /
• 2011

This paper introduces an automated method for building height recovery through the integration of high-resolution satellite images and digital vector maps. A cross-correlation matching method along the vertical line locus on the Ikonos images was deployed to recover building heights. The rational function models composed of rational polynomial coefficients were utilized to create a stereopair of the epipolar resampled Ikonos images. Building footprints from the digital maps were used for locating the vertical guideline along the building edges. The digital terrain model (DTM) was generated from the contour layer in the digital maps. The terrain height derived from the DTM at each foot of the buildings was used as the starting location for image matching. At a preset incremental value of height along the vertical guidelines derived from vertical line loci, an evaluation process that is based on the cross-correlation matching of the images was carried out to test if the top of the building has reached where maximum correlation occurs. The accuracy of the reconstructed buildings was evaluated by the comparison with manually digitized 3D building data derived from aerial photographs.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.13 no.9
• /
• pp.3856-3863
• /
• 2012

A 3D reconstruction technique from stereo images that requires minimal intervention from the user has been developed. The reconstruction problem consists of three steps of estimating specific geometry groups. The first step is estimating the epipolar geometry that exists between the stereo image pairs which includes feature matching in both images. The second is estimating the affine geometry, a process to find a special plane in the projective space by means of vanishing points. The third step, which includes camera self-calibration, is obtaining a metric geometry from which a 3D model of the scene could be obtained. The major advantage of this method is that the stereo images do not need to be calibrated for reconstruction. The results of camera calibration and reconstruction have shown the possibility of obtaining a 3D model directly from features in the images.