Abstract: A method for extracting point cloud feature includes: obtaining an original point cloud, conducting hybrid scale voxelization on the original point cloud; generating point-wise hybrid scale voxel features by feature encoding the point cloud subjected to the hybrid scale voxelization; and generating pseudo image feature maps using aggregated features and projection scale information. In this way, problems that at a single voxel scale, the inference time is longer when the voxel scale is smaller, and the larger voxel fails to capture intricate features and accurate location of smaller objects can be effectively overcome.

Abstract: An object detection network includes: a hybrid voxel feature extractor configured to acquire a raw point cloud, extract a hybrid scale voxel feature from the raw point cloud, and project the hybrid scale voxel feature to generate a pseudo-image feature map; a backbone network configured to perform a hybrid voxel scale feature fusion by using the pseudo-image feature map to generate multi-class pyramid features; and a detection head configured to predict a three-dimensional object box of a corresponding class according to the multi-class pyramid features. The object detection network can effectively solve a problem that under a single voxel scale, inference time is longer if the voxel scale is smaller, and an intricate feature cannot be captured and a smaller object cannot be accurately located if the voxel scale is larger. Different classes of 3D objects can be detected quickly and accurately in a 3D scene.

Abstract: A method for encoding a point cloud feature includes: obtaining an original point cloud, generating point-wise hybrid scale voxel features by conducting hybrid scale voxelization on the original point cloud; generating voxel-wise attention features using spatial features and voxel features of points in a voxel at each voxel scale; aggregating the hybrid scale voxel features and the attention features to obtain voxel-wise projection scale information; and mapping the voxel-wise projection scale information to pseudo image features at projection scales. In this way, points can be retained during voxelization, and a feature encoding network is guided by the attention features to be more attentive to interrelationships between points in the voxels, thereby increasing accuracy of three-dimensional object detection.

Abstract: An object detection network includes: a hybrid voxel feature extractor configured to acquire a raw point cloud, extract a hybrid scale voxel feature from the raw point cloud, and project the hybrid scale voxel feature to generate a pseudo-image feature map; a backbone network configured to perform a hybrid voxel scale feature fusion by using the pseudo-image feature map to generate multi-class pyramid features; and a detection head configured to predict a three-dimensional object box of a corresponding class according to the multi-class pyramid features. The object detection network can effectively solve a problem that under a single voxel scale, inference time is longer if the voxel scale is smaller, and an intricate feature cannot be captured and a smaller object cannot be accurately located if the voxel scale is larger. Different classes of 3D objects can be detected quickly and accurately in a 3D scene.

Abstract: A method for encoding a point cloud feature includes: obtaining an original point cloud, generating point-wise hybrid scale voxel features by conducting hybrid scale voxelization on the original point cloud; generating voxel-wise attention features using spatial features and voxel features of points in a voxel at each voxel scale; aggregating the hybrid scale voxel features and the attention features to obtain voxel-wise projection scale information; and mapping the voxel-wise projection scale information to pseudo image features at projection scales. In this way, points can be retained during voxelization, and a feature encoding network is guided by the attention features to be more attentive to interrelationships between points in the voxels, thereby increasing accuracy of three-dimensional object detection.

Abstract: A method for extracting point cloud feature includes: obtaining an original point cloud, conducting hybrid scale voxelization on the original point cloud; generating point-wise hybrid scale voxel features by feature encoding the point cloud subjected to the hybrid scale voxelization; and generating pseudo image feature maps using aggregated features and projection scale information. In this way, problems that at a single voxel scale, the inference time is longer when the voxel scale is smaller, and the larger voxel fails to capture intricate features and accurate location of smaller objects can be effectively overcome.