Browsing by Author "Chen, Can"

Now showing 1 - 4 of 4
  • Thumbnail Image
    ItemOpen Access
    GAPointNet: Graph attention based point neural network for exploiting local feature of point cloud
    (Elsevier, 2021-01-26) Chen, Can; Fragonara, Luca Zanotti; Tsourdos, Antonios
    Exploiting fine-grained semantic features on point cloud data is still challenging because of its irregular and sparse structure in a non-Euclidean space. In order to represent the local feature for each central point that is helpful towards better contextual learning, a max pooling operation is often used to highlight the most important feature in the local region. However, all other geometric local correlations between each central point and corresponding neighbourhood are ignored during the max pooling operation. To this end, the attention mechanism is promising in capturing node representation on graph-based data by attending over all the neighbouring nodes. In this paper, we propose a novel neural network for point cloud analysis, GAPointNet, which is able to learn local geometric representations by embedding graph attention mechanism within stacked Multi-Layer-Perceptron (MLP) layers. Specifically, we highlight different attention weights on the neighbourhood of each center point to efficiently exploit local features. We also combine attention features with local signature features generated by our attention pooling to fully extract local geometric structures and enhance the network robustness. The proposed GAPointNet architecture is tested on various benchmark datasets (i.e. ModelNet40, ShapeNet part, S3DIS, KITTI) and achieves state-of-the-art performance in both the shape classification and segmentation tasks
  • Thumbnail Image
    ItemOpen Access
    Go wider: an efficient neural network for point cloud analysis via group convolutions
    (MDPI, 2020-04-01) Chen, Can; Fragonara, Luca Zanotti; Tsourdos, Antonios
    In order to achieve a better performance for point cloud analysis, many researchers apply deep neural networks using stacked Multi-Layer-Perceptron (MLP) convolutions over an irregular point cloud. However, applying these dense MLP convolutions over a large amount of points (e.g., autonomous driving application) leads to limitations due to the computation and memory capabilities. To achieve higher performances but decrease the computational complexity, we propose a deep-wide neural network, named ShufflePointNet, which can exploit fine-grained local features, but also reduce redundancies using group convolution and channel shuffle operation. Unlike conventional operations that directly apply MLPs on the high-dimensional features of a point cloud, our model goes “wider” by splitting features into groups with smaller depth in advance, having the respective MLP computations applied only to a single group, which can significantly reduce complexity and computation. At the same time, we allow communication between groups by shuffling the feature channel to capture fine-grained features. We further discuss the multi-branch method for wider neural networks being also beneficial to feature extraction for point clouds. We present extensive experiments for shape classification tasks on a ModelNet40 dataset and semantic segmentation task on large scale datasets ShapeNet part, S3DIS and KITTI. Finally, we carry out an ablation study and compare our model to other state-of-the-art algorithms to show its efficiency in terms of complexity and accuracy
  • Thumbnail Image
    ItemOpen Access
    Relation3DMOT: exploiting deep affinity for 3D multi-object tracking from view aggregation
    (MDPI, 2021-03-17) Chen, Can; Zanotti Fragonara, Luca; Tsourdos, Antonios
    Autonomous systems need to localize and track surrounding objects in 3D space for safe motion planning. As a result, 3D multi-object tracking (MOT) plays a vital role in autonomous navigation. Most MOT methods use a tracking-by-detection pipeline, which includes both the object detection and data association tasks. However, many approaches detect objects in 2D RGB sequences for tracking, which lacks reliability when localizing objects in 3D space. Furthermore, it is still challenging to learn discriminative features for temporally consistent detection in different frames, and the affinity matrix is typically learned from independent object features without considering the feature interaction between detected objects in the different frames. To settle these problems, we first employ a joint feature extractor to fuse the appearance feature and the motion feature captured from 2D RGB images and 3D point clouds, and then we propose a novel convolutional operation, named RelationConv, to better exploit the correlation between each pair of objects in the adjacent frames and learn a deep affinity matrix for further data association. We finally provide extensive evaluation to reveal that our proposed model achieves state-of-the-art performance on the KITTI tracking benchmark.
  • Thumbnail Image
    ItemOpen Access
    RoIFusion: 3D object detection from LiDAR and vision
    (IEEE, 2021-04-01) Chen, Can; Zanotti Fragonara, Luca; Tsourdos, Antonios
    When localizing and detecting 3D objects for autonomous driving scenes, obtaining information from multiple sensors (e.g., camera, LIDAR) is capable of mutually offering useful complementary information to enhance the robustness of 3D detectors. In this paper, a deep neural network architecture, named RoIFusion, is proposed to efficiently fuse the multi-modality features for 3D object detection by leveraging the advantages of LIDAR and camera sensors. In order to achieve this task, instead of densely combining the point-wise feature of the point cloud with the related pixel features, our fusion method novelly aggregates a small set of 3D Region of Interests (RoIs) in the point clouds with the corresponding 2D RoIs in the images, which are beneficial for reducing the computation cost and avoiding the viewpoint misalignment during the feature aggregation from different sensors. Finally, Extensive experiments are performed on the KITTI 3D object detection challenging benchmark to show the effectiveness of our fusion method and demonstrate that our deep fusion approach achieves state-of-the-art performance