When Pedestrian Detection Meets Multi-Modal Learning: Generalist Model and Benchmark Dataset

Recent years have witnessed increasing research attention towards pedestriandetection by taking the advantages of different sensor modalities (e.g. RGB,IR, Depth, LiDAR and Event). However, designing a unified generalist model thatcan effectively process diverse sensor modalities remains a challenge. Thispaper introduces MMPedestron, a novel generalist model for multimodalperception. Unlike previous specialist models that only process one or a pairof specific modality inputs, MMPedestron is able to process multiple modalinputs and their dynamic combinations. The proposed approach comprises aunified encoder for modal representation and fusion and a general head forpedestrian detection. We introduce two extra learnable tokens, i.e. MAA andMAF, for adaptive multi-modal feature fusion. In addition, we construct theMMPD dataset, the first large-scale benchmark for multi-modal pedestriandetection. This benchmark incorporates existing public datasets and a newlycollected dataset called EventPed, covering a wide range of sensor modalitiesincluding RGB, IR, Depth, LiDAR, and Event data. With multi-modal jointtraining, our model achieves state-of-the-art performance on a wide range ofpedestrian detection benchmarks, surpassing leading models tailored forspecific sensor modality. For example, it achieves 71.1 AP on COCO-Persons and72.6 AP on LLVIP. Notably, our model achieves comparable performance to theInternImage-H model on CrowdHuman with 30x smaller parameters. Codes and dataare available at https://github.com/BubblyYi/MMPedestron.