In the facial expression recognition task, a good-performing convolutional neural network (CNN) model trained on one dataset (source dataset) usually performs poorly on another dataset (target dataset). This is because the feature distribution of the same emotion varies in different datasets. To improve the cross-dataset accuracy of the CNN model, we introduce an unsupervised domain adaptation method, which is especially suitable for unlabelled small target dataset. In order to solve the problem of lack of samples from the target dataset, we train a generative adversarial network (GAN) on the target dataset and use the GAN generated samples to fine-tune the model pretrained on the source dataset. In the process of fine-tuning, we give the unlabelled GAN generated samples distributed pseudolabels dynamically according to the current prediction probabilities. Our method can be easily applied to any existing convolutional neural networks (CNN). We demonstrate the effectiveness of our method on four facial expression recognition datasets with two CNN structures and obtain inspiring results.

The use of DMSO-doped poly (3,4-ethylenedioxythiophene):poly (4-styrenesulfonate) (PEDOT:PSS) thin films produced by spin coating as transparent electrodes is reported. The transmittance of these electrodes at THz frequencies, as well as their surface morphology and electrical conductivity, are subsequently investigated in relation to the thickness of their DMSO-doped PEDOT:PSS film. A maximum conductivity of 5078 S/cm is obtained for a single-layer film (52 nm) doped with 15 vol% DMSO in PEDOT:PSS solution, providing a transmittance of up to 83.5% at 1.22 THz. The newly proposed blue-phase liquid crystal THz modulator provides continuous tunability and full electrical controllability at THz frequencies.

A dynamic binocular stereo vision (DBSV) based on a non-zoom rotating camera in a large field of view (FOV) is established herein. A novel two-point method is proposed to estimate the initial parameters of the camera quickly. The intrinsic parameters and roll angle of each camera remains constant, and the pitch and yaw angles after the camera rotates are directly estimated from output of high-precision two-axes platform, which makes it possible that three-dimensional (3D) coordinate can be measured online after rotation. Moreover, a target matching algorithm based on moving DLT is proposed to achieve automatic alignment of the camera. The accuracy of 3D coordinate measurement is evaluated on various synthetic and real data, and the DBSV is suitable for occasions where extremely high accuracy is not required in large FOV.

Multi-camera systems are widely used in the fields of airborne remote sensing and unmanned aerial vehicle imaging. The measurement precision of these systems depends on the accuracy of the extrinsic parameters. Therefore, it is important to accurately calibrate the extrinsic parameters between the onboard cameras. Unlike conventional multi-camera calibration methods with a common field of view (FOV), multi-camera calibration without overlapping FOVs has certain difficulties. In this paper, we propose a calibration method for a multi-camera system without common FOVs, which is used on aero photogrammetry. First, the extrinsic parameters of any two cameras in a multi-camera system is calibrated, and the extrinsic matrix is optimized by the re-projection error. Then, the extrinsic parameters of each camera are unified to the system reference coordinate system by using the global optimization method. A simulation experiment and a physical verification experiment are designed for the theoretical arithmetic. The experimental results show that this method is operable. The rotation error angle of the camera’s extrinsic parameters is less than 0.001rad and the translation error is less than 0.08 mm.

Camera pose estimation from the image of a planar object has important applications in photogrammetry and computer vision. In this paper, an efficient approach to find the initial solutions for iterative camera pose estimation using coplanar points is proposed. Starting with homography, the proposed approach provides a least-squares solution for absolute orientation, which has a relatively high accuracy and can be easily refined into one optimal pose that locates local minima of the according error function by using Gauss-Newton scheme or Lu's orthogonal iteration algorithm. In response to ambiguities that exist in pose estimation from planar objects, we propose a novel method to find initial approximation of the second pose, which is different from existing methods in its concise form and clear geometric interpretation. Thorough testing on synthetic data shows that combined with currently employed iterative optimization algorithm, the two initial solutions proposed in this paper can achieve the same accuracy and robustness as the best state-of-the-art pose estimation algorithms, while with a significant decrease in computational cost. Real experiment is also employed to demonstrate its performance.

Nowadays, binocular stereo vision (BSV) is extensively used in real-time 3D reconstruction, which requires cameras to quickly implement self-calibration. At present, the camera parameters are typically estimated through iterative optimization. The calibration accuracy is high, but the process is time consuming. Hence, a system of BSV with rotating and non-zooming cameras is established in this study, in which the cameras can rotate horizontally and vertically. The cameras’ intrinsic parameters and initial position are estimated in advance by using Zhang’s calibration method. Only the yaw rotation angle in the horizontal direction and pitch in the vertical direction for each camera should be obtained during rotation. Therefore, we present a novel self-calibration method by using a single feature point and transform the imaging model of the pitch and yaw into a quadratic equation of the tangent value of the pitch. The closed-form solutions of the pitch and yaw can be obtained with known approximate values, which avoid the iterative convergence problem. Computer simulation and physical experiments prove the feasibility of the proposed method. Additionally, we compare the proposed method with Zhang’s method. Our experimental data indicate that the averages of the absolute errors of the Euler angles and translation vectors relative to the reference values are less than 0.21° and 6.6 mm, respectively, and the averages of the relative errors of 3D reconstruction coordinates do not exceed 4.2%.