BiologybioRxiv
Ranked by formulaVerdict: NotableDroneMegaDetector, A Generalised Model For Detecting Large Mammals In Aerial Imagery
Large-scale drone surveys can generate imagery across substantial geographic areas, but automated analysis remains challenging because animal appearance, target density, imaging conditions and target scale can vary…
Caveats
- Preprint; not yet peer reviewed.
VerdictWorth a reader's time today.
Abstract
Large-scale drone surveys can generate imagery across substantial geographic areas, but automated analysis remains challenging because animal appearance, target density, imaging conditions and target scale can vary considerably between surveys. We present a general-purpose RF-DETR Medium object detector developed for large-scale aerial drone imagery and evaluate its ability to generalise across taxa, environments and survey conditions. Unlike specialist counting systems that primarily return point detections, the proposed detector produces explicit bounding boxes, providing a spatial representation that can support counting as well as downstream detection, classification and tracking workflows. We evaluated the detector on an internal held-out test set and independent external datasets spanning cattle, birds, caribou and hippopotamuses, together with imagery acquired under substantially different survey conditions. On the comparative Cattle dataset, RF-DETR matched OWL across the reported point-localisation and counting metrics, while on Caribou it achieved almost identical point-based detection performance and similar aggregate counting error. Performance varied across the broader external benchmark, with reduced transfer to unseen avian imagery and differences between models across datasets. At high altitude on SAVMAP, all evaluated approaches experienced substantial performance degradation, demonstrating a shared limitation associated with small target scale. Overall, the results show that a general-purpose detector trained on heterogeneous drone imagery can provide competitive localisation and counting performance across diverse datasets while retaining explicit bounding-box outputs. The trained model weights and inference code are provided online to support reproducible large-scale aerial wildlife monitoring and further development of drone-based computer-vision workflows.
The editor's rubric
| Dimension | Level | Weight | What that level means |
|---|---|---|---|
| Leverage | ████░ 4 | 24% | A general-purpose tool used across several fields (Adam, ResNet, LoRA, next-generation sequencing). |
| Magnitude | ██░░░ 2 | 16% | Solid incremental gain on a meaningful problem. |
| Evidence | ███░░ 3 | 20% | Solid: multiple benchmarks or cohorts, ablations, fair baselines, released code or data. |
| Novelty | ██░░░ 2 | 20% | A new combination of known ideas. |
| Trajectory | ███░░ 3 | 10% | A clear path to scale. |
| Stakes | ██░░░ 2 | 10% | Benefits a professional community (practitioners, clinicians, engineers). |
Editor’s rationale
Heuristic triage from title and abstract text only, not a reading of the paper. Cues found: method (we propose); breadth (general-purpose, many tasks); verification (held-out test, independent replication); scale (scalable).
How the score was computed
- Merit
- 5.6 / 10
- Adjusted merit
- 4.7 / 10
- Attention
- 0%
- Freshness
- 88%