Multimodal Transformation for Small-Scale 3-D Object Detection in Industrial Scenarios

Detecting small and information-scarce objects within complex 3D backgrounds remains a critical yet challenging task in industrial scenarios. While existing multi-modal approaches leverage the cross-modal complementation between point clouds and intensity data to improve object representation, they still face limitations in addressing multi-modal fine alignment and achieving precise 3D positional regression. This paper proposes a new 3D object detection network based on feature fusion between point clouds and multi-view images. First, we propose a dense multi-modal feature fusion module that establishes point-to-pixel fine correspondence and effectively integrates multi-modal feature channels. Then, we design a normalized 3D positional embedding generator to enhance a transformer-based detection head, which improves localization accuracy through refined positional encoding of the fused features. Experimental results on a multi-modal industrial dataset demonstrate that the proposed method achieves state-of-the-art performance with an AP of 0.903 and a Recall of 94.10%, representing 6.35% and 3.58% improvements in comparison to the existing optimal method. Specifically, the method achieves optimal performance of 2.66 mm and 0.85 mm on the proposed metrics 3D mATE and 3D mASE, representing 19.64%, and 29.75% improvement over the existing optimal method, respectively.