Event

Release of a 3D Human Detection Dataset for Search and Rescue and a Three-Modality Fusion Method

1 reports · 1 sources

Markus Kramer et al. released TriSAR-3D, a three-modality human detection dataset, and proposed PaintFusion, a fusion method for 3D human detection in search and rescue scenarios. The dataset was collected using a 128-channel LiDAR, an RGB camera, and a thermal camera, indoors and outdoors and at different times of day, covering diverse poses including lying down. The dataset and code have been made public. The paper states that PaintFusion achieves 65.4 AP@50 on this dataset, higher than the LiDAR-only CenterPoint (53.4), the RGB+LiDAR BEVFusion (60.5), and the adapted three-modality Tri-BEVFusion (60.7). The paper also states that if lying-down samples are removed from the training data, the lying-pose AP@50 of all evaluated detectors drops to near zero.