NDNet

NDNet performs lossless multiscale spatial modeling to enable real-time semantic segmentation and high-resolution scene parsing for autonomous driving.

Key Features:

• Reversible ND and NC operations: Implements reversible ND and NC operations to alter channelwise information representation without information loss.
• Lossless resolution conversion: Preserves information integrity during resolution transformations through reversible operations.
• Complementary thumbnail sampling and collation: Uses complementary thumbnail sampling and collation to enhance spatial modeling capabilities.
• Local Capturer and Global Dependence Builder (LCGB): Performs initial processing for large-scale inputs with fast, lossless resolution reduction while extracting features enriched with global context.
• Spacewise Multiscale Feature Extractor (SMFE): Conducts dense multiscale spatial feature extraction efficiently with minimal computational overhead.
• High-Resolution Semantic Generator (HSG): Reconstructs high-resolution outputs and adaptively amends semantic confusions to produce precise segmentation.

Scientific Applications:

• Autonomous driving scene parsing: Applied to real-time semantic segmentation in autonomous driving to support safety-critical perception tasks.
• Benchmark evaluation on Cityscapes: Demonstrates performance on the Cityscapes dataset with reported 76.47% mean Intersection over Union (mIoU) at over 240 frames per second and 78.8% mIoU at over 150 frames per second.

Methodology:

NDNet leverages reversible ND and NC operations for lossless resolution transformations; LCGB preprocesses inputs with fast lossless reduction and global-context feature extraction; SMFE performs efficient multiscale dense feature extraction; HSG reconstructs high-resolution outputs and refines semantic details.