Toward Unified 3D Object Detection via Algorithm and Data Unification.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Unified 3D object detection is crucial for robot navigation but faces challenges due to diverse indoor/outdoor data characteristics.
• Training models on heterogeneous datasets with varying geometry and domain distributions leads to convergence instability.

Purpose of the Study:

• To develop algorithms and leverage data strategies for unified 3D object detection across diverse indoor and outdoor scenes.
• To address challenges in geometry learning ambiguity and domain distribution differences in 3D object detection models.

Main Methods:

• Proposed a two-stage monocular 3D object detector (UniMODE) using a bird's-eye-view (BEV) paradigm with an uneven BEV grid and sparse feature projection.
• Developed a unified domain alignment method to handle heterogeneous domains and incorporated depth information for multi-modal detection (MM-UniMODE).
• Introduced the first unified multi-modal 3D object detection benchmark, MM-Omni3D.

Main Results:

• Experimental results demonstrate the effectiveness of the proposed strategies, including the uneven BEV grid and sparse projection.
• The unified domain alignment method successfully handles heterogeneous domains, improving detection performance.
• Incorporating depth information significantly enhances training robustness and overall detection accuracy.

Conclusions:

• The proposed UniMODE and MM-UniMODE detectors effectively address challenges in unified 3D object detection.
• Multi-modal data, particularly depth information, offers significant benefits for improving 3D object detection robustness and performance.
• The MM-Omni3D benchmark provides a valuable resource for advancing multi-modal 3D object detection research.