M&M3D: Multi-Dataset Training and EfficientNetwork for Multi-view 3D Object Detection

Abstract The research focuses on the challenge of domain adaptation and visual data transfer in camera-only 3D object detection. Despite numerous advancements in borderline methods, their performance significantly decreases across domains. 'M\&M3D' is introduced as a novel training approach for multi-view 3D object detection utilizing exclusively camera data under the multi-dataset training. 'M\&M3D' leverages 2D multi-view image features, 3D anchor queries and 3D space information as positional information and relies on a modest amount of source data as the visual-language transfer design. 'M\&M3D' exhibits favorable performance on standard metrics of 3D object detection with a smaller network size, distinguishing itself from dataset-specific models that require longer training time and individual fine-tuning on source domain. It might ignite future deployment of models on smaller devices.