ImVoxelGNet: Image to voxels geometry-aware projection for multi-view RGB-based 3D object detection

Area of Science:

• Computer Vision
• 3D Object Detection

Background:

• 3D object detection from images is challenging due to integrating geometric perception.
• Existing methods inadequately utilize pixel features during voxel-pixel alignment, reducing accuracy.

Purpose of the Study:

• To propose ImVoxelGNet, a novel network framework for enhanced 3D object detection.
• To improve geometric perception and scene understanding in multi-view 3D object detection.

Main Methods:

• ImVoxelGNet integrates pixel features using an expansion operation to enhance spatial geometric learning.
• An implicit geometric perception structure refines features and learns voxel occupancy relationships.
• Final predictions are generated using a detection head with 3D convolutions.

Main Results:

• ImVoxelGNet achieved up to a 2.2% improvement in mean average precision (mAP) on the ScanNetV2 dataset.
• The method demonstrates significant enhancement in 3D object detection performance.

Conclusions:

• The proposed ImVoxelGNet effectively improves 3D object detection by enhancing geometric perception.
• Comprehensive scene understanding is achieved through better integration of visual and geometric data.