Food recognition is a key component in evaluation of everyday food intakes, and its challenge is due to intraclass variation. In this paper, we present an automatic food classification method, DietCam, which specifically addresses the variation of food appearances. DietCam consists of two major components, ingredient detection and food classification. Food ingredients are detected through a combination of a deformable part-based model and a texture verification model. From the detected ingredients, food categories are classified using a multiview multikernel SVM. In the experiment, DietCam presents reliability and outperformance in recognition of food with complex ingredients on a database including 15,262 food images of 55 food types.

This paper proposes the recognition framework of car makes and models from a single image captured by a traffic camera. Due to various configurations of traffic cameras, a traffic image may be captured in different viewpoints and lighting conditions, and the image quality varies in resolution and color depth. In the framework, cars are first detected using a part-based detector, and license plates and headlamps are detected as cardinal anchor points to rectify projective distortion. Car features are extracted, normalized, and classified using an ensemble of neural-network classifiers. In the experiment, the performance of the proposed method is evaluated on a data set of practical traffic images. The results prove the effectiveness of the proposed method in vehicle detection and model recognition.

This paper proposes an ego-motion tracking method that utilizes visual-inertial sensors for wearable blind navigation. The unique challenge of wearable motion tracking is to cope with arbitrary body motion and complex environmental dynamics. We introduce a visual sanity check to select accurate visual estimations by comparing visually estimated rotation with measured rotation by a gyroscope. The movement trajectory is recovered through adaptive fusion of visual estimations and inertial measurements, where the visual estimation outputs motion transformation between consecutive image captures, and inertial sensors measure translational acceleration and angular velocities. The frame rates of visual and inertial sensors are different, and vary with respect to time owning to visual sanity checks. We hence employ a multirate extended Kalman filter (EKF) to fuse visual and inertial estimations. The proposed method was tested in different indoor environments, and the results show its effectiveness and accuracy in ego-motion tracking.

Depth sensing is an important problem in many applications such as autonomous driving, robotics, and automation. This paper presents an adaptive active fusion method for scene depth estimation by using an RGB camera and a single-point LiDAR. An active scanning mechanism is proposed to guide laser scanning based on critical visual and saliency features, and the convolutional spatial propagation network (CSPN) is designed to generate and refine full depth map from the sparse depth scans. The active scanning mechanism generates a depth mask by using log-spectrum saliency detection, Canny edge detection, and uniform sampling, which indicate critical regions that require a high resolution of laser scanning. To reconstruct a full depth map, the designed CSPN network extracts affinity matrices from the sparse depth scans, while reserving global spatial information in the RGB images. The performance of proposed method was evaluated and compared with the state-of-the-art methods on the NYUv2 dataset, and the experiment demonstrated its outperformance in reconstruction accuracy and robustness to measurement noise. The proposed method was also evaluated in real-world scenarios.

Given the description of an object, s physical attributes, humans can determine a proper strategy and grasp an object. This paper proposes an approach to determine grasping strategy for an anthropomorphic robotic hand simply based on natural-language descriptions of an object. A learning-based approach is proposed to help a robotic hand learn suitable grasp poses starting from the natural language description of the object. Object features are parsed from natural-language descriptions by using a customized natural-language processing technique. The most likely grasp type for the given object is learned from the human grasping taxonomy based on the parsed features. The grasping strategy generated by the proposed approach is evaluated both by simulation study and execution of the grasps on an AR10 robotic hand.