Beyond Euclidean Tokens: Hyperbolic Structure-Aware Mapping for Dual-Task Scene Parsing With Only Minimal Trainable...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unified scene parsing for semantic segmentation and depth estimation is crucial for real-world perception but challenging.
• Existing methods struggle with domain shifts, leading to degraded semantic segmentation.
• Hierarchical calibration gaps in Euclidean representations limit preserving semantic structures.

Purpose of the Study:

• To develop a novel framework, HyperMapper, for unified cross-domain scene parsing.
• To address the limitations of Euclidean representations in preserving semantic hierarchies.
• To achieve robust semantic segmentation and depth estimation across diverse domains without retraining.

Main Methods:

• Proposed HyperMapper, a hyperbolic structure-aware mapping framework.
• Utilized hyperbolic token-to-feature interactions to bridge semantic understanding and geometric priors.
• Combined vision foundation models (VFMs) with parameter-efficient fine-tuning (PEFT) for cross-domain adaptation.

Main Results:

• HyperMapper demonstrated improved performance in semantic segmentation (mIoU) for both parent and child categories.
• The framework maintained strong depth estimation capabilities without retraining.
• Achieved consistent segmentation accuracy improvements over strong baselines across multiple benchmarks.

Conclusions:

• HyperMapper offers a promising direction for task-preserving dual-task adaptation in scene parsing.
• The hyperbolic approach effectively bridges semantic and geometric learning.
• Established a pathway toward unified, cross-domain scene parsing capabilities.