Two research achievements from Xiamen University's School of Informatics have been accepted by IEEE Transactions on Pattern Analysis and Machine Intelligence, a leading international journal in artificial intelligence and computer science, as well as a Class-A journal recommended by the China Computer Federation.
In the first study, the researchers focused on multi-agent collaborative perception, a key technology for autonomous driving, robotic collaboration, and low-altitude flight. As the number of agents increases, training collaborative detectors typically requires substantial manual annotation. To address this challenge, the team previously proposed an unsupervised method for Detecting Objects from Multi-Agent LiDAR scans, known as DOtA, which generates preliminary labels using internally shared information from collaborative agents.
Building on this framework, the new method DOtA++ enhances performance by introducing composite prior constraints, including multi-agent observation consistency and point cloud geometric distribution constraints. Experiments show that DOtA++ achieves a 10.7 percent improvement in mean Average Precision over traditional unsupervised methods on the V2X-R dataset.
The second study addresses the heavy reliance of conventional 3D object detectors on extensive human annotations. The team introduced a Commonsense Prototype-based Detector (CPD) for unsupervised 3D object detection, which constructs commonsense prototypes to represent objects' geometric centers and sizes and uses them to generate high-quality pseudo-labels.
Building on this framework, CPD++ further leverages motion cues, learning localization from stationary objects and recognition from moving objects to enable knowledge transfer between them. When trained on the Waymo Open Dataset and tested on KITTI, CPD++ achieves 89.25 percent 3D Average Precision for the moderate car class at a 0.5 IoU threshold, reaching 95.3 percent of the performance of fully supervised methods.
