Search

In the last few years, the availability of temporal knowledge graphs (TKGs), which associate time information for each event, increased the need for completing in these TKGs. In order to solve the completion problem, the existing efforts extend static knowledge graph completion models to handle time-dependent representations, or encode the structural information of events for passing temporal message. However, the above efforts mostly construct TKGs in the form of multi-edges mesh and focus on entity features. They do not explore the power of relations for TKG completion task. In this paper, we introduce a form of relational multi-chains to reconstruct TKGs and propose a relation-oriented attention mechanism for embedding the features of relations. According to our attention mechanism, we build a Relation-oriented Attention Network (RoAN) to model temporal embeddings of relations. It is worth noting that our approach is model-agnostic and can be potentially combined with most existing TKG completion models. Experimental results show that our approach can be coupled with previous TKGC methods and can increase their performance accordingly. In addition, the analysis reveals the principle of our relation-oriented attention mechanism. [ABSTRACT FROM AUTHOR]

Lane detection is a fundamental task in urban traffic management. Like lane detection for automatic driving, lane detection for traffic management will be faced with common challenges including fog, night, shadow, no line and etc. Meanwhile, it will be faced with new and unique challenges including variable number of lanes, blocked lanes due to oversize vehicles and color interference. In this paper, we propose a multi-lane detection method by combining the line anchor and feature shift (MLD-LAFS) to cope with these challenges. The proposed method features a two-branch neural network structure based on the line anchor. The first branch is the feature shift branch incorporating spatial attention, which is designed to enhance the local features of lanes. The second branch is the global information branch incorporating channel attention and cross attention, which is designed to establish the long-distance connection of lane features. The channel attention can obtain the global channel information. The uncoupled feature shift cross attention can obtain the global spatial information. The feature map containing the global information can be obtained by fusing the feature maps of the first and second branches. The line anchor is used as the supervision information to generate the predicted lane and realize multi-lane detection. The feature shift is used to solve the problems of lane line being blocked and color interference. We perform the performance evaluation on three datasets, including CULane, TuSimple and a newly constructed dataset MonitorLane. Experimental results show that the proposed MLD-LAFS achieves remarkable results on the CULane and the TuSimple dataset. Moreover, the proposed MLD-LAFS achieves the highest grading in F1-score on the MonitorLane dataset, compared to existing solutions, including LaneATT, PolyLaneNet, Lane Shape Prediction with Transformers (LSTR) and etc. [ABSTRACT FROM AUTHOR]