Enhancing Transparent Object Pose Estimation: A Fusion of GDR-Net and Edge Detection

TL;DR

This work tackles RGB-only 6D pose estimation of transparent objects by introducing edge-detection as a preprocessing step to amplify boundary cues. The authors integrate edge representations with a GDR-Net–based pose estimator and a YOLOX detector, evaluating four configurations on the synthetic Trans6D-32K dataset. Results show that Canny edge preprocessing can improve mean $ADD$-S scores and pose accuracy for several objects, though edge augmentation does not consistently benefit object detection. The study highlights the potential of edge representations to aid challenging transparent-object perception and points to future work on broader datasets and textured objects to generalize the findings.

Abstract

Object pose estimation of transparent objects remains a challenging task in the field of robot vision due to the immense influence of lighting, background, and reflections. However, the edges of clear objects have the highest contrast, which leads to stable and prominent features. We propose a novel approach by incorporating edge detection in a pre-processing step for the tasks of object detection and object pose estimation. We conducted experiments to investigate the effect of edge detectors on transparent objects. We examine the performance of the state-of-the-art 6D object pose estimation pipeline GDR-Net and the object detector YOLOX when applying different edge detectors as pre-processing steps (i.e., Canny edge detection with and without color information, and holistically-nested edges (HED)). We evaluate the physically-based rendered dataset Trans6D-32 K of transparent objects with parameters proposed by the BOP Challenge. Our results indicate that applying edge detection as a pre-processing enhances performance for certain objects.

Figures (3)

• Figure 1: Experimental setup. We conduct four sets of experiments: In the first experiment, the Trans6D-32K dataset without any augmentation (vanilla) is trained and tested. The second experiment investigates the performance when Canny is applied before training. The third experiment involves HED as an pre-processing step. Finally, we conducted a set of experiments incorporating color information and canny edges.
• Figure 2: Simplified GDR-Net framework and method. Given an RGB image $X$, we introduced an optional pre-processing step in which the edge detector is applied and the augmented image is generated. Then, we train the object detector from which zoomed-in RoI are retrieved. This serves as input for the geometric feature regression consisting of the trained GDR-Net and a PnP algorithm to directly regress the 6D object pose estimate.
• Figure 3: Dataset augmentation. (a) Unaugmented sample images of the Trans6d-32 K dataset (b) Sample images after having applied the Canny edge detector. (c) Sample images after having applied HED. (d) Sample images providing canny edges and color information