Research map / Detection and segmentation / Open-set and few-shot detection
WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WildlifeMapper, a SAM-based transformer detector with a high-frequency feature generator and query refiner for multi-species wildlife detection in aerial imagery, and releases the Mara-Wildlife dataset.
Problem
Automated detection, localization and identification of multiple wildlife species in aerial survey imagery, which generalizes poorly across species and landscapes due to homogeneous backgrounds and weak local structure capture.
Approach
Suppresses low-frequency background via a Fourier high-pass branch (High-Frequency Feature Generator) whose output refines both image features (Feature Refiner) and learnable queries (Query Refiner) in a single-stage DETR-style detector built on a SAM ViT encoder, with a large-kernel dilated patch embedding for context.
Key results
- WildlifeMapper reaches 0.56 mAP on Mara-Wildlife versus 0.30 for the best baseline (YOLOv5).
- On the merged Virunga-Garamba-AED datasets, WildlifeMapper achieves 0.68 mAP versus 0.48 for YOLOv5 and YOLOv8.
- Ablation shows HFG in both FR and QR raises mAP from 0.46 to 0.56 on Mara-Wildlife.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HFG and FR to the model (ablation) | WildlifeMapper · Mara-Wildlife | mAP: 0.46 -> 0.54 | Helped |
| add HFG to QR while omitting FR (ablation) | WildlifeMapper · Mara-Wildlife | mAP: 0.46 -> 0.49 | Helped |
| use HFG with both FR and QR (ablation) | WildlifeMapper · Mara-Wildlife | mAP: 0.46 -> 0.56 | Helped |
| increase the kernel size from 7 × 7 to 16 × 16 (ablation) | WildlifeMapper · Mara-Wildlife | mAP: 0.55 -> 0.558 | Helped |
| increase the kernel size from 16 × 16 to 32 × 32 (ablation) | WildlifeMapper · Mara-Wildlife | mAP: 0.558 -> 0.57 | Helped |
| provide HFG output to QR using cross-attention (ablation) | WildlifeMapper · Mara-Wildlife | mAP: +0.03mAP | Helped |
| merge HFG and image features by cross-attention rather than addition or concatenation (ablation) | WildlifeMapper · Mara-Wildlife | mAP | Helped |
| replace cross-attention in the feature merge with addition or concatenation (ablation) | WildlifeMapper · Mara-Wildlife | mAP | Hurt |
| train and test across geographic regions, with training only on Kenya | WildlifeMapper · four common species across Kenya and other regions | mAP | Mixed |
| train and test on the tern domain | WildlifeMapper · tern | accuracy | Mixed |
Surprising findings
- Standard detectors trained on the aerial datasets appear to learn homogeneous background features rather than object features, per the authors' explanation for poor baseline performance.
Limitations
- Accuracy drops when animals are clustered in shadows, such as inside bomas.
- Rocks, small trees, and artifacts are sometimes detected as animals (false positives).
- Dataset covers only 21 large mammal species from one ecosystem; baseline results on public datasets were reproduced by the authors, not verified against original code.
Experimental setup
Datasets: Mara-Wildlife Virunga Garamba AED tern bird dataset
Benchmarks: Mara-Wildlife Virunga-Garamba-AED
Metrics: mAP mAP50 mIoU counting MAE
Baselines: Faster R-CNN DETR Co-DETR YOLOv5 YOLOv8 LSKNet DroneDetect
Builds on: Segment Anything Model DETR Vision Transformer
Closest papers by meaning
- BirdCollect: A Comprehensive Benchmark for Analyzing Dense Bird Flock Attributes (AAAI 2024 · similarity 0.88)
- BuckTales: A multi-UAV dataset for multi-object tracking and re-identification of wild antelopes (NeurIPS 2024 · similarity 0.86)
- WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife (NeurIPS 2026 · similarity 0.86)
- RealBirdID: Benchmarking Bird Species Identification in the Era of MLLMs (CVPR 2026 · similarity 0.84)
- Combining Observational Data and Language for Species Range Estimation (NeurIPS 2024 · similarity 0.84)
- Open-Insect: Benchmarking Open-Set Recognition of Novel Species in Biodiversity Monitoring (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.