Interpretable Instance Segmentation for Camouflaged Wildlife Detection

What the paper does

Owais Ansari co-authored a peer-reviewed segmentation framework for detecting camouflaged wildlife, presented at ADCIS 2025, where it won Best Presentation. The work combines YOLOv8n-seg instance segmentation with two explainable AI techniques — Seg-Grad-CAM and LIME — so that an ecological monitoring system can justify each detection rather than only report one.

Camouflaged animals are the adversarial case for segmentation: the object and its background share texture, colour and edge statistics, which is precisely the signal most architectures rely on. A model can reach a respectable score on this task while keying on artefacts that have nothing to do with the animal. For ecological monitoring, where detections inform conservation decisions, a mask without a reason is difficult to trust.

Why interpretability was the point

The framework treats interpretability as a first-class output rather than a diagnostic afterthought. Seg-Grad-CAM extends gradient-based class activation mapping to segmentation, so the explanation is spatially aligned with the predicted mask — it shows which pixels drove the mask, not merely which drove a label. LIME contributes a model-agnostic perturbation view, approximating local behaviour by occluding superpixels and measuring the response.

Using both matters because they fail differently. Gradient methods inherit the model’s own biases; perturbation methods are model-agnostic but sensitive to how regions are segmented in the first place. Agreement between the two is meaningfully stronger evidence than either alone.

Benchmarking

Segmentation accuracy and interpretability were benchmarked against U-Net and FCN baselines — established encoder–decoder architectures for dense prediction. Testing a lightweight real-time architecture against them frames the practical question directly: how much accuracy does a field-deployable model give up, and is what remains explainable enough to act on?