EfficientPose 6D: Scalable and Efficient 6D Object Pose Estimation

TL;DR

This work targets real-time 6D object pose estimation by enlarging the GDRNPP-based family with 40 candidate architectures and introducing the Adaptive Margin-Dependent Iterative Selection (AMIS) algorithm to automatically balance inference time and pose accuracy across diverse datasets. AMIS identifies robust sweet spots in the inference time–accuracy space, enabling model selection tailored to application constraints. Empirical results show that AMIS-selected models can deliver up to 35% faster inference with minimal accuracy loss, or about 25% accuracy gains with modest time increases, across LM-O, YCB-V, T-LESS, and ITODD. The approach offers a practical path to fast, scalable 6D pose estimation suitable for industrial real-time feedback and robotic manipulation, with avenues for further optimization in temporal fusion and end-to-end lightweight designs.

Abstract

In industrial applications requiring real-time feedback, such as quality control and robotic manipulation, the demand for high-speed and accurate pose estimation remains critical. Despite advances improving speed and accuracy in pose estimation, finding a balance between computational efficiency and accuracy poses significant challenges in dynamic environments. Most current algorithms lack scalability in estimation time, especially for diverse datasets, and the state-of-the-art (SOTA) methods are often too slow. This study focuses on developing a fast and scalable set of pose estimators based on GDRNPP to meet or exceed current benchmarks in accuracy and robustness, particularly addressing the efficiency-accuracy trade-off essential in real-time scenarios. We propose the AMIS algorithm to tailor the utilized model according to an application-specific trade-off between inference time and accuracy. We further show the effectiveness of the AMIS-based model choice on four prominent benchmark datasets (LM-O, YCB-V, T-LESS, and ITODD).

Figures (7)

• Figure 1: Architecture of GDRNPP / GDR-Net wang2021gdr
• Figure 2: Relative required runtime of the six conceptual stages during inference of GDR-Net / GDRNPP.
• Figure 3: Intermediate results during backbone selection using k-means clustering and the Successive Halving Algorithm with the mean of MSPD, MSSD and VSD as selection criterion.
• Figure 4: Geo Head Variations GDR-Net architecture with the adaptations Geo Head variation 1 and 2, which reduce the number of up-sample layer of the vanilla Geo Head part.
• Figure 5: Candidates for adding connection in vanilla Geo Head, Geo Head variation 1, Geo Head variation 2. The connection can be added in one of positions labeled ①, ② and ③.