Contour-based instance segmentation method of road scene.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Instance segmentation is vital for autonomous driving, but current contour-based methods struggle with inaccurate initial contours.
• Errors in initial contours propagate, hindering subsequent deformation modules and increasing learning complexity.

Purpose of the Study:

• To propose novel methods for accurate and efficient instance segmentation in autonomous driving scenarios.
• To address the limitations of existing contour-based methods in handling complex road scene information.

Main Methods:

• Introduced Contour Initialization based on Instance Center Features (CIICF) for precise initial contour prediction.
• Developed Contour Deformation based on Differentiation Module (CDDM) using 1D convolution for inherent positional learning.
• Incorporated absolute position encoding to enhance model positional awareness.

Main Results:

• CIICF and CDDM significantly improved final contour accuracy across multiple datasets (Cityscapes, KINS, SBD).
• Performance gains observed: 6.9%, 5.1%, and 9.1% increases in final contour metrics.
• Enhanced contour generation speed from 43.2 FPS to 49.8 FPS compared to baseline.

Conclusions:

• The proposed CIICF and CDDM methods offer superior performance for instance segmentation in autonomous driving.
• These advancements contribute to more robust and efficient perception systems for self-driving vehicles.
• The improved accuracy and speed demonstrate the potential for real-world application.