MonoLite3D: Lightweight 3D Object Properties Estimation

TL;DR

This work targets efficient monocular estimation of 3D object properties for autonomous perception on resource-limited hardware. It introduces MonoLite3D, a one-stage Orientation-Dimensions Estimator that leverages a GhostNet backbone and a MultiBin orientation head to predict object orientation and dimensions from cropped 2D boxes, achieving strong KITTI orientation scores with real-time inference on embedded GPUs. The approach emphasizes lightweight design (5.61M parameters) and a training setup on the KITTI dataset with augmentation and simple preprocessing, reporting competitive accuracy (e.g., Easy/Moderate/Hard orientation: $90.23\%$, $82.27\%$, $69.81\%$) and fast inference times (as low as $0.0121$ s for batch of 200). Overall, MonoLite3D demonstrates that high-quality 3D pose and size estimation can be achieved from monocular inputs in real-time on constrained hardware, enabling practical deployment in autonomous vehicles and robotics.

Abstract

Reliable perception of the environment plays a crucial role in enabling efficient self-driving vehicles. Therefore, the perception system necessitates the acquisition of comprehensive 3D data regarding the surrounding objects within a specific time constrain, including their dimensions, spatial location and orientation. Deep learning has gained significant popularity in perception systems, enabling the conversion of image features captured by a camera into meaningful semantic information. This research paper introduces MonoLite3D network, an embedded-device friendly lightweight deep learning methodology designed for hardware environments with limited resources. MonoLite3D network is a cutting-edge technique that focuses on estimating multiple properties of 3D objects, encompassing their dimensions and spatial orientation, solely from monocular images. This approach is specifically designed to meet the requirements of resource-constrained environments, making it highly suitable for deployment on devices with limited computational capabilities. The experimental results validate the accuracy and efficiency of the proposed approach on the orientation benchmark of the KITTI dataset. It achieves an impressive score of 82.27% on the moderate class and 69.81% on the hard class, while still meeting the real-time requirements.

Figures (5)

• Figure 1: 3D object detection deep learning methodologies.
• Figure 2: The general architecture of the proposed MonoLite3D network is composed of one stage, which is the Orientation-Dimensions Estimator stage.
• Figure 3: The object's orientation, denoted as $\theta$, is computed by adding $\theta_{ray}$ and $\theta_l$ together. The Orientation-Dimensions Estimator provides the value for $\theta_l$, while $\theta_{ray}$ can be determined in relation to the center of the object's bounding box using the known camera intrinsic parameters.
• Figure 4: Orientation-Dimensions Estimation heads.
• Figure 5: The general architecture of the proposed MonoLite3D network is composed of one stage, which is the Orientation-Dimensions Estimator stage.