FFAM: Feature Factorization Activation Map for Explanation of 3D Detectors

TL;DR

This paper introduces FFAM, a novel method employing non-negative matrix factorization to generate high-quality, interpretable visual explanations for LiDAR-based 3D object detectors, addressing the black-box nature of current models.

Contribution

The paper proposes FFAM, a new explanation technique for 3D detectors that produces concept activation maps and object-specific explanations, filling a gap in interpretability for LiDAR-based models.

Findings

FFAM produces high-quality visual explanations for 3D detectors.

Experimental results demonstrate FFAM's effectiveness across multiple datasets.

FFAM outperforms existing explanation methods in clarity and relevance.

Abstract

LiDAR-based 3D object detection has made impressive progress recently, yet most existing models are black-box, lacking interpretability. Previous explanation approaches primarily focus on analyzing image-based models and are not readily applicable to LiDAR-based 3D detectors. In this paper, we propose a feature factorization activation map (FFAM) to generate high-quality visual explanations for 3D detectors. FFAM employs non-negative matrix factorization to generate concept activation maps and subsequently aggregates these maps to obtain a global visual explanation. To achieve object-specific visual explanations, we refine the global visual explanation using the feature gradient of a target object. Additionally, we introduce a voxel upsampling strategy to align the scale between the activation map and input point cloud. We qualitatively and quantitatively analyze FFAM with multiple detectors on several datasets. Experimental results validate the high-quality visual explanations produced by FFAM. The Code will be available at \url{https://github.com/Say2L/FFAM.git}.