Semantic Consistency Reasoning for 3-D Object Detection in Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• 3-D object detection from point clouds is crucial for autonomous systems.
• Existing methods primarily focus on geometric features, overlooking semantic properties of points.
• Inconsistency between geometric and semantic cues limits detection performance.

Purpose of the Study:

• To propose a novel semantic consistency (SC) mechanism for 3-D object detection.
• To enhance the understanding of semantic relations between 3-D object boxes and internal points.
• To improve the robustness and accuracy of 3-D object detection models.

Main Methods:

• Developed a semantic consistency (SC) mechanism based on the principle that object box semantics should match internal point semantics.
• Proposed the SC network (SCNet) comprising feature extraction, detection decision, and semantic segmentation modules.
• Jointly trained the semantic segmentation module with detection modules to refine model parameters.

Main Results:

• The SC mechanism significantly boosts performance in 3-D object detection.
• SCNet demonstrated improved accuracy on ScanNetV2, SUN RGB-D, and KITTI datasets.
• The SC mechanism proved model-agnostic, enhancing three different 3-D object detectors.

Conclusions:

• The proposed semantic consistency mechanism effectively integrates semantic and geometric information for 3-D object detection.
• SCNet offers a robust and accurate solution for detecting objects in point clouds.
• The SC mechanism's versatility allows for broad application across various 3-D detection frameworks.