Detecting 6D Poses of Target Objects From Cluttered Scenes by Learning to Align the Point Cloud Patches With the CAD Models

Abstract

6D target object detection is of great importance to many applications such as robotics, industrial automation, and unmanned vehicles and is increasingly influencing broad industries including manufacturing, transportation, and retail industries, to name a few. Unlike the more common object detection methods that use the two-dimensional data such as RGB or depth images, the method proposed here relies on the three-dimensional point cloud of target objects to detect them in cluttered scenes in an end-to-end fashion. However, conventional point cloud-based 6D object detection methods rely on the robustness of key-point detection results that are not straightforward for humans to understand. The drawback makes conventional point cloud-based methods require expert knowledge to tune. In this paper, a point cloud-based 6D target object detection method that uses segmented object point cloud patches to predict object 6D poses and identity is introduced. In this way, the key-point detection is substituted with a point cloud segmentation procedure that is easier to visualize and tune. To extract the object pose and identity information from point cloud patches, we proposed an end-to-end data-driven pose correction model that can be trained with synthetic data and run on real-world data. A simple yet efficient basis spanning layer booster is proposed to accelerate the learning process and improve the pose estimation precision. An alignment error-based loss function is also introduced to make the proposed pose correction model trainable, efficiently. Experiments show that although the proposed model is trained only using object CAD models, its 6D detection performance matches that of the models using view data. Thus, the proposed method is suitable for 6D detection applications that have object CAD models instead of labeled scene data.

Deep Learning on Monocular Object Pose Detection and Tracking: A Comprehensive Overview

“…Since only point cloud is required to be processed, the network architecture proposed by Gao et al [77] is much more lightweight than methods that take both RGB data and depth data as input. Then, Chen et al [96] use segmented object point cloud patches from the target object's CAD model as input and form a pose prediction model without the need for fine-tuning, thus saving computational cost. Additionally, G2L-Net [78] decouples the prediction pipeline into global localization, translation localization, and rotation localization and estimates object pose through a coarse-to-fine manner without using a CAD model, achieving real-time performance.…”

Intelligent Analysis of Computer-Aided Garment Design CAD Based on Virtual Reality Model for Surplus Fabric Removal and Reuse without Segmentation of Cutting Pieces

“…e top and bottom can be designed by the user or the design stored in the system. In either case, it should be a reasonable stacking of the top and bottom [11].…”

Weakly-Supervised Object Detection Learning through Human-Robot Interaction

“…Latest research on object detection for robotics mainly focuses on improving precision in particularly difficult scenarios such as, occlusion and clutter [3], [4], [5], [6], [7]. To this aim, a widely used approach is to rely on deep learning based methods (like e.g.…”