Det-LIME: Detector-Aware, Multi-Instance Local Interpretable Model-Agnostic Explanations for Automated Marine Mammal Detection
Jiayi Zhou, David W. Johnston, Brinnae Bent
cs.CV · cs.AI
Quick verdict
Det-LIME extends LIME to multi-instance detection with large attribution gains on seal and penguin tasks; promising, but no ablations or detector baselines.
Summary
The paper introduces Det-LIME, a detector-aware extension of LIME for object detection in ecological imagery. Motivated by the observation that classification-oriented explainers explain one image-level score and blur evidence across multiple animals, the method (i) segments images with constrained SLIC superpixels, (ii) weights all detections of a target class (confidence-, area-, or uniform-weighted), (iii) scores perturbations via per-instance IoU-matched detection confidence (capped at 0.95) with an IoU-dependent bonus, and (iv) converts superpixel coefficients into a pixel map using box-overlap and a Gaussian proximity kernel centered on each predicted box. On harbor seal detection with Faster R-CNN (76 test images) and penguin detection with YOLOv9c (321 test images), Mean Attribution Ratio rises from at most 25% to 48.8% (seals) and from 4.7% to 39.3% (penguins); Max Saliency Hit Rate rises from 31.3% to 92.4% and from 4.4% to 82.2%. An adjusted one-to-one comparison controlling for multi- vs single-instance outputs yields similar or larger gains. Failure-case analyses (black-ice false positives; boundary and albatross confusions) illustrate diagnostic value. Main limitations: no ablation isolating components, hyperparameters chosen from pilot runs, no quantitative comparison with purpose-built detector explainers (D-RISE, SoDeX), and evaluation metrics that reward the box-proximity prior built into the method.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Conventional LIME and its stabilized variants cannot provide instance-level explanations for multi-object detection because they explain a single image-level score with coarse, globally weighted perturbations.
- Combining per-instance weighting, IoU-based instance matching across perturbations, and box-centered spatial focusing yields instance-specific, box-aligned explanations using only black-box detector outputs.
- Det-LIME substantially outperforms LIME, DLIME, and SLIME on Attribution Ratio and Max Saliency Hit Rate for both harbor seal (Faster R-CNN) and penguin (YOLOv9c) detection.
- Instance-aware attribution reveals failure modes such as black-ice false positives and habitat-boundary errors, enabling targeted augmentation, threshold setting, and more defensible conservation decisions.
Strengths
Well-motivated, clearly scoped gap
Sections 1-2 convincingly argue that classification-oriented LIME explains an image-level score with coarse perturbations that engulf multiple animals, a real problem for colonial/social species monitoring where many detections share a frame.
Large and consistent quantitative gains
Table 1: Max Saliency Hit Rate improves from 31.34% (best baseline, DLIME) to 92.42% for seals and from 4.38% to 82.20% for penguins; Mean Attribution Ratio roughly doubles or better in both tasks, with the adjusted one-to-one comparison showing the same pattern.
Thoughtful two-tier evaluation protocol
Model-agnostic design validated across architectures
Honest failure-case discussion
Weaknesses
No ablation study
Metric-method circularity
Missing detection-specific baselines
Unjustified and partly unspecified hyperparameters
Thin statistical reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Det-LIME?
Det-LIME is a detector-aware adaptation of LIME that explains each bounding-box detection in an image separately. It uses superpixel perturbations, per-instance weighting, IoU-based matching of the same instance across perturbations, and box-centered spatial focusing to produce pixel-level, instance-specific attribution maps for object detectors without access to model internals.
How much better is Det-LIME than standard LIME?
On harbor seal drone imagery with Faster R-CNN, mean Attribution Ratio rose from 16-25% for LIME variants to 48.8%, and Max Saliency Hit Rate from at most 31.3% to 92.4%. On penguin imagery with YOLOv9c, Attribution Ratio rose from 4.7% or less to 39.3% and hit rate from 4.4% or less to 82.2%.
Does Det-LIME require access to the neural network's internals?
No. It needs only the input image and detector outputs (bounding boxes, class labels, confidence scores), so it works with third-party or deployed detection systems, unlike gradient-based methods such as Grad-CAM or LayerCAM which require feature maps and gradients.
What are the main limitations of Det-LIME?
The paper reports no ablation isolating each component, several hyperparameters were chosen from pilot runs, the evaluation metrics measure alignment with bounding boxes which favors the method's box-centered design, superpixel segmentation adds parameter sensitivity, and runtime cost and field usability were not assessed.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 could not be assessed from the extracted text; qualitative claims about heatmap quality rest on the authors' descriptions.
- The metadata abstract is truncated mid-sentence, but the full in-paper abstract is present in the extracted text.
- Some hyperparameter values (IoU match threshold, perturbation count, SLIC settings) are absent from the text and may exist only in the released code.
- Table 1 values were transcribed from linearized text; minor formatting risk cannot be excluded.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.