FastPoseCNN: Real-Time Monocular Category-Level Pose and Size Estimation Framework

TL;DR

FastPoseCNN introduces a real-time, efficient framework for monocular category-level pose and size estimation from a single RGB image, significantly improving inference speed over previous methods.

Contribution

The paper presents a novel, fast framework using ResNet-FPN and decoupled decoders for pose and size estimation, enabling real-time performance for unseen objects.

Findings

Achieves real-time inference at 25+ fps

Outperforms previous methods in accuracy and speed

Performs global context estimation for multiple objects

Abstract

The primary focus of this paper is the development of a framework for pose and size estimation of unseen objects given a single RGB image - all in real-time. In 2019, the first category-level pose and size estimation framework was proposed alongside two novel datasets called CAMERA and REAL. However, current methodologies are restricted from practical use because of its long inference time (2-4 fps). Their approach's inference had significant delays because they used the computationally expensive MaskedRCNN framework and Umeyama algorithm. To optimize our method and yield real-time results, our framework uses the efficient ResNet-FPN framework alongside decoupling the translation, rotation, and size regression problem by using distinct decoders. Moreover, our methodology performs pose and size estimation in a global context - i.e., estimating the involved parameters of all captured objects in the image all at once. We perform extensive testing to fully compare the performance in terms of precision and speed to demonstrate the capability of our method.