Multi-granularity knowledge distillation and prototype consistency regularization for class-incremental learning.

Deep neural networks (DNNs) are prone to the notorious catastrophic forgetting problem when learning new tasks incrementally. Class-incremental learning (CIL) is a promising solution to tackle the challenge and learn new classes while not forgetting old ones. Existing CIL approaches adopted stored representative exemplars or complex generative models to achieve good performance. However, storing data from previous tasks causes memory or privacy issues, and the training of generative models is unstable and inefficient. This paper proposes a method based on multi-granularity knowledge distillation and prototype consistency regularization (MDPCR) that performs well even when the previous training data is unavailable. First, we propose to design knowledge distillation losses in the deep feature space to constrain the incremental model trained on the new data. Thereby, multi-granularity is captured from three aspects: by distilling multi-scale self-attentive features, the feature similarity probability, and global features to maximize the retention of previous knowledge, effectively alleviating catastrophic forgetting. Conversely, we preserve the prototype of each old class and employ prototype consistency regularization (PCR) to ensure that the old prototypes and semantically enhanced prototypes produce consistent prediction, which excels in enhancing the robustness of old prototypes and reduces the classification bias. Extensive experiments on three CIL benchmark datasets confirm that MDPCR performs significantly better over exemplar-free methods and outperforms typical exemplar-based approaches.

Exact and Heuristic Multi-Robot Dubins Coverage Path Planning for Known Environments.

Coverage path planning (CPP) of multiple Dubins robots has been extensively applied in aerial monitoring, marine exploration, and search and rescue. Existing multi-robot coverage path planning (MCPP) research use exact or heuristic algorithms to address coverage applications. However, several exact algorithms always provide precise area division rather than coverage paths, and heuristic methods face the challenge of balancing accuracy and complexity. This paper focuses on the Dubins MCPP problem of known environments. Firstly, we present an exact Dubins multi-robot coverage path planning (EDM) algorithm based on mixed linear integer programming (MILP). The EDM algorithm searches the entire solution space to obtain the shortest Dubins coverage path. Secondly, a heuristic approximate credit-based Dubins multi-robot coverage path planning (CDM) algorithm is presented, which utilizes the credit model to balance tasks among robots and a tree partition strategy to reduce complexity. Comparison experiments with other exact and approximate algorithms demonstrate that EDM provides the least coverage time in small scenes, and CDM produces a shorter coverage time and less computation time in large scenes. Feasibility experiments demonstrate the applicability of EDM and CDM to a high-fidelity fixed-wing unmanned aerial vehicle (UAV) model.

ESVIO: Event-Based Stereo Visual-Inertial Odometry.

The emerging event cameras are bio-inspired sensors that can output pixel-level brightness changes at extremely high rates, and event-based visual-inertial odometry (VIO) is widely studied and used in autonomous robots. In this paper, we propose an event-based stereo VIO system, namely ESVIO. Firstly, we present a novel direct event-based VIO method, which fuses events' depth, Time-Surface images, and pre-integrated inertial measurement to estimate the camera motion and inertial measurement unit (IMU) biases in a sliding window non-linear optimization framework, effectively improving the state estimation accuracy and robustness. Secondly, we design an event-inertia semi-joint initialization method, through two steps of event-only initialization and event-inertia initial optimization, to rapidly and accurately solve the initialization parameters of the VIO system, thereby further improving the state estimation accuracy. Based on these two methods, we implement the ESVIO system and evaluate the effectiveness and robustness of ESVIO on various public datasets. The experimental results show that ESVIO achieves good performance in both accuracy and robustness when compared with other state-of-the-art event-based VIO and stereo visual odometry (VO) systems, and, at the same time, with no compromise to real-time performance.