Vehicle Re-identification (Re-ID) refers to finding the same vehicle shot by other cameras from a given vehicle image library, which can also be regarded as a sub-problem of image retrieval. It plays an important role in intelligent transportation and smart cities. The key of vehicle Re-ID is to extract discriminative vehicle features. To better extract such features from the vehicle image to improve the recognition accuracy, we propose a three-branch adaptive attention network-Global Relational Attention and Multi-granularity Feature Learning (GRMF) to improve feature representation and discrimination. First, we divide the network into three branches, extracting different and useful features from three perspectives: spatial location, channel information, and local information. Second, we propose two effective global relational attention modules, which capture the global structural information for better attention learning. Specifically, to determine the importance level of a node, we use the global relationship between the node and all other nodes to infer the attention weight of the node directly. Third, according to the characteristics of the vehicle re-identification task, we introduce a suitable local partition strategy. It not only can simply capture subtle local information but also solve the problem of misalignment and within-part consistency disruption to a great extent. Extensive experiments demonstrate the effectiveness of our approach, and we achieve state-of-the-art results on two challenging datasets, including VeRi776 and VehicleID.

Direction relation is an important spatial relation. Descriptions and representations for direction relations have different levels of detail because of the varying dimensions of spatial objects and different scales of the embedding spaces.Based on a direction-relation matrix, the hierarchical frame of spatial direction relations which partitions direction relations orderly and thoroughly is built. Interior direction relations are used to perfect the representation of direction relations and the binary-encoding idea is creatively applied to construct an interior detailed matrix describing multiple interior direction relations by a uniform matrix. The model integrates topological information into the description model for direction relations, which will lay the foundations of spatial compositive reasoning.

No abstract

No abstract

The Gao Fen Xiang Ji (GFXJ) is the first Chinese self–developed airborne three–line array charge-coupled devices (CCD) camera and is designed to meet 8 cm ground sample distance (GSD), 0.5 m planimetry accuracy, and 0.28 m elevation accuracy for ground three-dimensional (3D) points at a flight height of 2000 m. These values also meet the 1:1000 scale mapping requirements in China. However, the original direct geopositioning accuracy of the GFXJ is approximately 4 m in the planimetry direction and 6 m in the elevation direction. To meet the ground 3D point accuracy requirements and improve the direct geopositioning accuracy of the GFXJ, this paper carries out a deep investigation on the GFXJ geometric calibration. This geometric calibration includes two main parts: the Global Navigation Satellite System (GNSS) lever arms and inertial measurement unit (IMU) boresight misalignment calibration, and the camera lens and CCD line distortion calibration. First, a brief introduction is given on the imaging properties of the GFXJ camera. Then, the GNSS lever arms and IMU boresight misalignment calibration models are built for the GFXJ camera. Next, a piecewise self-calibration model based on the CCD viewing angle is established for the GFXJ lens and CCD line distortion calibration. Subsequently, an iterative two-step calibration scheme is proposed for the geometric calibration. Finally, experiments were implemented using multiple flight blocks obtained in the Songshan remote sensing comprehensive field and the Hegang area of Heilongjiang Province. Through calibration experiments, geometric calibration values were obtained for the GNSS lever arms and IMU boresight misalignment. Reliable CAM files were independently generated for the forward, nadir, and backward line arrays. The experiments showed that the proposed GNSS lever arms and IMU boresight misalignment calibration models and the piecewise self-calibration model had good applicability and effectiveness for the GFXJ camera. The proposed two-step calibration scheme can significantly enhance the geometric positioning accuracy of the GFXJ camera. The original direct geopositioning accuracy of the GFXJ is approximately 4 m in the planimetry direction and 6 m in the elevation direction. Using the GNSS lever arms and the IMU boresight misalignment calibration values and the CAM files, the positioning accuracy of the GFXJ camera can fulfill the 3D point accuracy requirements and the 1:1000 mapping accuracy requirements at a 2000 m flight height after aerial triangulation with only several ground control points. The planimetry accuracy is approximately 0.2 m, and the elevation accuracy is less than 0.28 m. In addition, the calibration models and calibration scheme established in this paper can provide a reference for calibration studies on other airborne linear array CCD cameras.