Xiaobin Liu

Many RGB-D simultaneous localization and mapping (SLAM) systems solely rely on point features, suffering from performance degradation in weak-texture scenes where line features can be complementary. However, existing RGB-D SLAM methods fusing point and line features commonly fail to explicitly estimate the feature uncertainties, hindering an accurate evaluation of both reliabilities and contributions of different features. In this paper, an RGB-D SLAM approach termed PLUE-SLAM integrating the closed-form uncertainty estimation of point and line features is proposed. Specifically, PLUE-SLAM derives the uncertainties of point and line features by jointly considering image observations, depth measurements and triangulation. In the front-end tracking, feature uncertainties provide an adaptive weighting mechanism for feature fusion, and are further propagated to the uncertainties of the camera poses. The uncertainties of both features and poses are used to weight the cost function in local bundle adjustment (BA) optimization, and are subsequently updated. Finally, a loop closure detection method using the feature uncertainties is designed to improve the robustness of PLUE-SLAM. Extensive experiments on public datasets and in real-world environments demonstrate that PLUE-SLAM outperforms state-of-the-art visual SLAM systems. PLUE-SLAM is open-sourced to benefit the community¹.

Vehicle Re-Identification (Re-ID) is challenging because vehicles of the same model commonly show similar appearance. We tackle this challenge by proposing a Global-Regional Feature (GRF) that depicts extra local details to enhance discrimination power in addition to the global context. It is motivated by the observation that, vehicles of same color, maker, and model can be distinguished by their regional difference, e.g., the decorations on the windshields. To accelerate the GRF learning and promote its discrimination power, we propose a Group-Group Loss (GGL) to optimize the distance within and across vehicle image groups. Different from the siamese or triplet loss, GGL is directly computed on image groups rather than individual sample pairs or triplets. By avoiding traversing numerous sample combinations, GGL makes the model training easier and more efficient. Those two contributions highlight this work from previous methods on vehicle Re-ID task, which commonly learn global features with triplet loss or its variants. We evaluate our methods on two large-scale vehicle Re-ID datasets, i.e., VeRi and VehicleID. Experimental results show our methods achieve promising performance in comparison with recent works.

The laborious manual person ID annotation results in limited training data and increased difficulty in learning discriminative representations. Meanwhile, high dimensional deep features are not ready for fast indexing and matching. Those challenges hinder the application of person Re-Identification (ReID) in large-scale data. To conquer those challenges, we propose a novel training strategy to learn compact binary hash codes. To facilitate feature learning, person images are decomposed into body parts, which are then composed across images into new positive and negative training samples. Binary code quality restrictions are also applied the during training procedure. Requiring no extra annotation costs, our algorithm iteratively generates hard training samples by itself and makes discriminative hash code learning with a limited number of labeled data possible. We hence use "self-guided" to describe this training procedure. Extensive experiments are conducted on two large-scale person ReID datasets, i.e., Market1501 and MSMT17 with distractors, showing our method is competitive compared with recent works.