Deep learning of directional truncated signed distance function for robust 3D object recognition

Abstract

In this paper, we develop a novel 3D object recognition algorithm to perform detection and pose estimation jointly. We focus on analyzing the advantages of the 3D point cloud relative to the RGB-D image and try to eliminate the unpredictability of output values that inevitably occurs in regression tasks. To achieve this, we first adopt the Truncated Signed Distance Function (TSDF) to encode the point cloud and extract low compact discriminative feature via unsupervised deep learning network. This approach can not only eliminate the dense scale sampling for offline model training but also reduce the distortion by mapping the 3D shape to the 2D plane and overcome the dependence on color cues. Then, we train a Hough forests to achieve multi-object detection and 6-DoF pose estimation simultaneously. In addition, we propose a robust multilevel verification strategy that effectively reduces the unpredictability of output values which occurs in the hough regression module. Experiments on public datasets demonstrate that our approach provides effective results comparable to the state-of-the-arts.

Efficient 3D object recognition via geometric information preservation

“…4.1. Results on the LC-HF dataset [25,27] This dataset [25,27] The statistic recognition results are shown in Table 1 , where the overall average F1-score of our method is 93.9%, in comparison with LineMod (74.0%) [21] , LC-HF (65.1%) [25] , ConvAE (74.7%) [27] , VoxelAE (76.8%) [28] and SSD-6D (88.5%) [29] respectively. Here, [29] counts a detection to be correct when the IoU score of a predicted bounding box with the groundtruth box is higher than 0.5.…”

“…Instead, W. Kehl et al [27] train a Convolution Autoencoder to extract patch features and estimate 6-DOF pose based on K-nn search, which gives better performance. Liu et al [28] present a 3D Voxel Autoencoder by converting the point clouds into voxel grids for fully using the 3D spatial structure information.…”

“…The decoders have the same structure to the encoders, which are used to reconstruct the input point clouds features. Several papers have proved that the features extracted by unsupervised reconstruction have effective performances [26][27][28] . In our case, the network has three stacked encoder layers and three symmetric decoder layers.…”

“…While most existing patch feature-based methods [25][26][27] could extract various features from RGB-D images to solve such problems, they inevitably involve mapping the 3D real world into 2D image space and induce the loss of 3D spatial information accordingly. In comparison, [28] tries to handle this problem by converting the point clouds into regular voxel grids and extracting 3D patch features accordingly. However, this introduces unnecessary computational cost.…”

“…Recently, a variety of CNN-based methods [29][30][31] have tried to learn features from large amounts of identically distributed training data, which depends heavily on large volume of data. The foregoing issues motivate us to extract a more efficient 3D features representation from raw point clouds directly and estimate the 6-DOF pose of the 3D objects depending on the hypotheses generation and verification strategy as shown in Generally, there are three main stages involved for 3D object recognition: 1) 3D feature extraction: Unlike the existing methods [25][26][27][28] and motivated by [32,33] , we design a 3D feature encoder, which enables point interaction within a local neighborhood sphere, by stacking the locally sphere-wise aggregated feature on point-wise features. Stacking multiple encoder layers allows further learning deep sphere-wise features and point-wise features.…”