Low-light enhancement task is an essential component of computer low-level visual tasks, which involves processing images captured under dim lighting conditions to make them appear as if they were taken under normal i...
详细信息
Limited statistical frame number and strong backscatter interference from smoke result in a photon-starved regime, severely limiting the depth imaging capability of array Gm-APD lidar in smoky environment. Here, we pr...
详细信息
This paper proposes an error analysis method based on epipolar geometry which is analyzed for improving the accuracy of attitude estimation. The influence of equivalent focal length and principal point coordinates on ...
详细信息
This paper addresses the problem of maneuvering multi-target tracking by a network of sensors having different and limited fields of view (FoV s). Each local sensor runs the Gaussian Mixture Probability Hypothetical D...
详细信息
Reasonable control organizational structure can help an unmanned system cluster cooperate more effectively to complete tasks. Previous research of existing organizational structures has problems implementing the task ...
详细信息
In this paper, a new power allocation scheme of the distributed radar system is proposed for mini-UAV tracking tasks in urban environments, considering the influence of the building occlusion on the probability of det...
详细信息
In order to monitor the health status of the rudder, this paper proposes a health monitoring method based on deep neural network through feature extraction. First, sufficient samples are obtained by data preprocessing...
详细信息
In this paper, based on the fusion tracking of infrared and radar, a federated kalman filtering algorithm based on adaptive fault tolerance is designed. According to the observation equation of infrared and radar, a d...
详细信息
Multi-target Tracking (MTT) is the process of processing received measurements to maintain estimates of the current status of multiple targets, with important applications to autonomous driving, aerial reconnaissance,...
Multi-target Tracking (MTT) is the process of processing received measurements to maintain estimates of the current status of multiple targets, with important applications to autonomous driving, aerial reconnaissance, underwater operations, and others. In the model-based setting, Bayesian filtering can provide the theoretical optimal estimate in a single target scenario. However, in complex situations, uncertain factors such as changes in the number of targets will cause the amount of calculation to increase exponentially, resulting in a decline in tracking accuracy. To solve that problem, model-free methods based on deep-learning provide an attractive alternative, especially the state-of-the-art architecture Transformer based encoder-decoder prediction model, which outperforms the Bayesian filters in the single frame prediction tasks. However, when switching to continuous tracking, these algorithms need to be trained separately frame by frame to adapt to the new tasks. Still, there is no correlation between their predictions from different frames, which prevents them from fully utilizing all the measurements. In this paper, we propose an end-to-end Transformer based MTT method with state autoregression, which allows the model to have the capability of online continuous tracking and make total use of the entire trajectory. The results show that the proposed model is a great extension from single-frame prediction to online continuous tracking.
Zero-shot object detection aims to localize and recognize objects of unseen classes. Most of existing works face two problems: the low recall of RPN in unseen classes and the confusion of unseen classes with backgroun...
Zero-shot object detection aims to localize and recognize objects of unseen classes. Most of existing works face two problems: the low recall of RPN in unseen classes and the confusion of unseen classes with background. In this paper, we present the first method that combines DETR and meta-learning to perform zero-shot object detection, named Meta-ZSDETR, where model training is formalized as an individual episode based meta-learning task. Different from Faster R-CNN based methods that firstly generate class-agnostic proposals, and then classify them with visual-semantic alignment module, Meta-ZSDETR directly predict class-specific boxes with class-specific queries and further filter them with the predicted accuracy from classification head. The model is optimized with meta-contrastive learning, which contains a regression head to generate the coordinates of class-specific boxes, a classification head to predict the accuracy of generated boxes, and a contrastive head that utilizes the proposed contrastive-reconstruction loss to further separate different classes in visual space. We conduct extensive experiments on two benchmark datasets MS COCO and PASCAL VOC. Experimental results show that our method outperforms the existing ZSD methods by a large margin.
暂无评论