Det-LIME: Detector-Aware, Multi-Instance Local Interpretable Model-Agnostic Explanations for Automated Marine Mammal Detection

Problem

This paper addresses a significant gap in explainability techniques for black-box object detectors, particularly in the context of marine mammal research. Existing methods lack the capability to provide interpretable insights into the decisions made by these models, which is crucial for understanding and improving detection performance in ecological studies. The work is presented as a preprint and has not yet undergone peer review.

Method

The authors propose Det-LIME, a detector-aware, multi-instance adaptation of the Local Interpretable Model-agnostic Explanations (LIME) framework. The core algorithm integrates several innovative components: 1) per-detection weighting to enhance the relevance of explanations, 2) a proximity kernel to better capture the spatial relationships in the data, and 3) Intersection-over-Union (IoU)-based matching to align explanations with detected instances. The method is applied to aerial drone imagery specifically for harbor seal detection, with an additional case study involving seabird detection. The training compute requirements are not specified in the paper.

Results

Det-LIME demonstrates improved performance over several baseline methods, including vanilla LIME, Stabilized LIME, Deterministic LIME, and gradient-based attribution methods. Key metrics reported include an improved Attribution Ratio and Max Saliency Hit Rate, although specific numerical values are not provided in the available text.

Limitations

The authors do not report any limitations in their work, and no obvious limitations are identified in the provided text.

Why it matters

The implications of this research are significant for downstream work in ecological monitoring and conservation efforts. By enhancing the explainability of object detection models, researchers can gain deeper insights into model behavior, leading to better-informed decisions in marine mammal conservation strategies. This work sets a precedent for future studies aiming to bridge the gap between model interpretability and practical ecological applications.