Unsupervised 3D Object Detection by Commonsense Clue.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Traditional 3D object detection methods heavily rely on extensive human annotations.
• This reliance poses a significant bottleneck for developing robust 3D detectors.

Purpose of the Study:

• To develop an unsupervised 3D object detector that does not require human annotations.
• To improve the performance and efficiency of 3D object detection through novel unsupervised techniques.

Main Methods:

• Proposed Commonsense Prototype-based Detector (CPD) using Commonsense Prototypes (CProto) for object representation.
• Developed CPD++ by incorporating motion cues, learning localization from stationary objects and recognition from moving objects.
• Utilized size and geometry priors from CProto to generate pseudo-labels and guide detector convergence.

Main Results:

• CPD and CPD++ outperform existing state-of-the-art unsupervised 3D object detectors.
• CPD++ achieved 89.25% 3D Average Precision (AP) on the moderate car class (0.5 IoU) on KITTI dataset, trained on Waymo Open Dataset (WOD).
• CPD++ reached 95.3% of the performance of fully supervised methods.

Conclusions:

• The proposed unsupervised methods, CPD and CPD++ significantly advance 3D object detection capabilities.
• Leveraging commonsense prototypes and motion cues offers a promising direction for annotation-free 3D object detection.
• These findings demonstrate the potential to achieve high performance in 3D object detection without manual annotation efforts.