ABSTRACT
This paper proposes MS-DISC, an extension to DISC that integrates hierarchical feature fusion into its distance-weighted extraction framework, aiming to capture multi-scale semantic information crucial for robust scene understanding in unstructured environments.
PAPER · PDF
Loading PDF...
Key findings
MS-DISC introduces a Scale-Space Feature Pyramid module for multi-scale feature extraction.
A Hierarchical Distance-Weighted Fusion mechanism aggregates features across scales.
A Cross-Scale Quality Gating module selects optimal scale representations based on view quality metrics.
Expected to achieve 8-12% improvement in mAcc for small object detection while maintaining real-time performance.
Limitations & open questions
The paper is a research proposal and thus does not include experimental results or conclusions.