Search

Semantic segmentation, as a pixel-level recognition task, has been widely used in a variety of practical scenes. Most of the existing methods try to improve the performance of the network by fusing the information of high and low layers. This kind of simple concatenation or element-wise addition will lead to the problem of unbalanced fusion and low utilization of inter-level features. To solve this problem, we propose the Inter-Level Feature Balanced Fusion Network (IFBFNet) to guide the inter-level feature fusion towards a more balanced and effective direction. Our overall network architecture is based on the encoder–decoder architecture. In the encoder, we use a relatively deep convolution network to extract rich semantic information. In the decoder, skip-connections are added to connect and fuse low-level spatial features to restore a clearer boundary expression gradually. We add an inter-level feature balanced fusion module to each skip connection. Additionally, to better capture the boundary information, we added a shallower spatial information stream to supplement more spatial information details. Experiments have proved the effectiveness of our module. Our IFBFNet achieved a competitive performance on the Cityscapes dataset with only finely annotated data used for training and has been greatly improved on the baseline network. [ABSTRACT FROM AUTHOR]

3D object detection in LiDAR point clouds has been extensively used in autonomous driving, intelligent robotics, and augmented reality. Although the one-stage 3D detector has satisfactory training and inference speed, there are still some performance problems due to insufficient utilization of bird's eye view (BEV) information. In this paper, a new backbone network is proposed to complete the cross-layer fusion of multi-scale BEV feature maps, which makes full use of various information for detection. Specifically, our proposed backbone network can be divided into a coarse branch and a fine branch. In the coarse branch, we use the pyramidal feature hierarchy (PFH) to generate multi-scale BEV feature maps, which retain the advantages of different levels and serves as the input of the fine branch. In the fine branch, our proposed pyramid splitting and aggregation (PSA) module deeply integrates different levels of multi-scale feature maps, thereby improving the expressive ability of the final features. Extensive experiments on the challenging KITTI-3D benchmark show that our method has better performance in both 3D and BEV object detection compared with some previous state-of-the-art methods. Experimental results with average precision (AP) prove the effectiveness of our network. [ABSTRACT FROM AUTHOR]

Spatiotemporal graph convolution has made significant progress in skeleton-based action recognition in recent years. Most of the existing graph convolution methods take all the joints of the human skeleton as the overall modeling graph, ignoring the differences in the movement patterns of various parts of the human, and cannot well connect the relationship between the different parts of the human skeleton. To capture the unique features of different parts of human skeleton data and the correlation of different parts, we propose two new graph convolution methods: the whole graph convolution network (WGCN) and the part graph convolution network (PGCN). WGCN learns the whole scale skeleton spatiotemporal features according to the movement patterns and physical structure of the human skeleton. PGCN divides the human skeleton graph into several subgraphs to learn the part scale spatiotemporal features. Moreover, we propose an adaptive fusion module that combines the two features for multiple complementary adaptive fusion to obtain more effective skeleton features. By coupling these proposals, we build a whole and part adaptive fusion graph convolution neural network (WPGCN) that outperforms previous state-of-the-art methods on three large-scale datasets: NTU RGB+D 60, NTU RGB+D 120, and Kinetics Skeleton 400. [ABSTRACT FROM AUTHOR]

Three-dimensional object detection from point cloud data is becoming more and more significant, especially for autonomous driving applications. However, it is difficult for lidar to obtain the complete structure of an object in a real scene due to its scanning characteristics. Although the existing methods have made great progress, most of them ignore the prior information of object structure, such as symmetry. So, in this paper, we use the symmetry of the object to complete the missing part in the point cloud and then detect it. Specifically, we propose a two-stage detection framework. In the first stage, we adopt an encoder–decoder structure to generate the symmetry points of the foreground points and make the symmetry points and the non-empty voxel centers form an enhanced point cloud. In the second stage, the enhanced point cloud is input into the baseline, which is an anchor-based region proposal network, to generate the detection results. Extensive experiments on the challenging KITTI benchmark show the effectiveness of our method, which has better performance on both 3D and BEV (bird's eye view) object detection compared with some previous state-of-the-art methods. [ABSTRACT FROM AUTHOR]