论文采用了基于voxel的方法(voxel feature encoding (VFE) module),先将点云数据划分成voxel,然后用VFE网络提取特征。每个voxel就相当于transformer中的一个token。考虑到transformer充足的感受野,和户外检测物体的尺寸都比较小,论文采用single-stride的网络设计。不会减少feature
The voxel size is held as (0.32m, 0.32m, 0.1875m) and we downsample the voxel feature map using our attention-style 3D pooling module only in the Z-Axis with the stride {4, 4, 2}. The window sizes along the Z-Axis are {32, 8, 2, 1}, and the maximum numbe...
In this section, a dynamic fine-tuning stacked auto-encoder neural network is demonstrated. First, the problems related to dynamic data environment are addressed. Second, the model of feature extraction and stacked auto-encoder neural-network in DDDAS is introduced. Third, the proposed system is ...
为此,作者计算了每个视频帧的低分辨率(low-resolution)3D占用体(occupancy volume),每个体素(voxel)都存储一个二进制值,该二进制值用来指示体素是否被占用。占用体是从学习的场景表示中提取的,当评估的密度高于阈值时,我们将体素标记为已占用。由于占用量具有低分辨率并且以二进制格式存储,因此300帧视频的占用量仅占用...
Lastly, the voxel selection in [66] is applied at the end of feature learning and thus has only a mild impact on latency. In addition, [10] proposes to select a subset of voxels as in- put to sparse convolutions. An important difference be- tween our...
Finally, we apply outlier removal and voxel grid filtering to reduce the point count. Through the above process, the algorithm we proposed can very well remove the afterimages generated by dynamic objects in the point cloud map, as already shown in Section 4.3. 4. Experimental evaluation The ex...
Structured light method is one of the best methods for automated 3D measurement in industrial production due to its stability and speed. However, when the surface of industrial parts has high dynamic range (HDR) areas, e.g. rust, oil stains, or shiny surfaces, phase calculation errors may ha...
The voxel feature fvoxeli ∈ Rh∗l is the hidden feature of neural voxels. and we utilize bilinear interpolation to interpolate the four nearby queried voxel features. 体素特征 fvoxi ∈ Rh∗l 是神经体素的隐藏特征。我们利用双线性插值对四个附近的查询体素特征进行插值。 In the compact represent...
T1-weighted MRI data was acquired using a 3D magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence (repetition time [TR] = 2400 ms; echo time [TE] = 2.14 ms; field of view [FOV] = 224 × 224 mm2; voxel size = 0.7 mm isotropic; number of slices = 256). T2-...
The storage (1740) stores instructions for the software (1780) implementing one or more of the disclosed innovations for point cloud compression with motion compensation and/or filtering to help correct voxel distortion errors, and corresponding decompression. The input device(s) (1750) may be a ...