Point AP @ 2.25 m = 0.9002 on 439 held-out full-resolution test images (2,970 annotated elephants). Best F1 0.8896 at confidence 0.50 (precision 0.9168 / recall 0.8640).


Base model	`Ultralytics/YOLO26` → `yolo26s.pt`
Relation	Full fine-tune (all 902/902 layers transferred from COCO)
Parameters	9,948,638 (9.95 M) · 19.5 MB checkpoint
Input	1024 × 1024 tiles, via SAHI over full-resolution photos
Classes	1 (`0: elephant`)
Head	`end2end: True` — YOLO26 is NMS-free (`reg_max: 1`)

The -p2 and -p6 variants were tried and rejected: they ship as architecture-only YAML, so .load('yolo26s.pt') transfers only 360/902 layers and leaves the neck and heads training from scratch. On 17k tiles that is a bad trade for the extra small-object resolution.

⚠️ Read this before you run it

Source images are ~5472×3648 and this model was trained on 1024 px tiles. Calling model.predict() on a full-resolution photo yields recall of 0.1%. You must use tiled inference (SAHI). Working example below.

Usage

pip install ultralytics sahi huggingface_hub

python

from huggingface_hub import hf_hub_download
from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

weights = hf_hub_download("iamhelitha/elefind-yolo26s", "elefind_yolo26s.pt")

model = AutoDetectionModel.from_pretrained(
    model_type="ultralytics",
    model_path=weights,
    confidence_threshold=0.30,
    device="cuda:0",          # "cpu" works, ~14 s per 20 MP image
)

result = get_sliced_prediction(
    "your_aerial_photo.jpg",
    model,
    slice_height=1024,
    slice_width=1024,
    overlap_height_ratio=0.30,
    overlap_width_ratio=0.30,
    # Pass these EXPLICITLY. SAHI defaults to GREEDYNMM/IOS, which merges
    # cross-tile duplicates instead of deleting them. Because this model emits
    # altitude-scaled boxes, IOS (intersection-over-smaller) over-merges animals
    # standing close together in herds and costs 0.047 point AP.
    postprocess_type="NMS",
    postprocess_match_metric="IOU",
    postprocess_match_threshold=0.40,
)

for obj in result.object_prediction_list:
    print(obj.bbox.to_xyxy(), obj.score.value)

Recommended settings: slice=1024, overlap=0.30, conf=0.30, iou=0.40. For maximum F1 use conf=0.50.

Latency (~20 MP image, RTX 3060): GPU ≈ 3.2 s, CPU ≈ 14.4 s. This is fixed by tiling geometry, not by how many elephants are present — an empty image costs the same as one with 62 animals.

The thing you need to understand about the metrics

AED ships point annotations only — image, x, y. There is no ground-truth bounding box anywhere in the dataset. Every box used to train or score any model on AED is invented, which means IoU-based mAP measures the assumed animal size at least as much as it measures the detector.

The bars above are the same detections from two models, graded three ways. Point matching calls them tied. The box metrics disagree with each other by 0.31 and then 0.74 — driven entirely by whether you assume an elephant is 3.5 m or 4.5 m across. Nobody measured that number.

So the primary metric here is point-based: a detection is correct if its centre falls within a tolerance expressed in metres of an annotated point, converted to pixels per image using that image's Ground Sampling Distance. No invented box, GSD-invariant, and it matches AED's own benchmark.

Evaluation results

Protocol


Test set	439 full-resolution AED test images (438 with annotations)
Ground truth	2,970 annotated elephant points
Inference	SAHI, 1024 px tiles, 0.30 overlap, match threshold 0.40
Capture floor	conf 0.05, so thresholds can be swept post-hoc
Postprocess	`NMS` / `IOU` (both reported; SAHI's default differs — see below)
AP rule	101-point rectangle interpolation (precision = 0 beyond max recall)
Primary metric	point-distance match, tolerance in metres, converted per image via GSD
Hardware	RTX 3060 12 GB, ~3.2 s per 20 MP image

The test split is fully held out — 438 source photos from AED's test_images, never seen in training or validation, and disjoint at source-photo level from both.

Reproduce with the evaluation code:

bash

python evaluation/evaluate.py \
    --model elefind_yolo26s=elefind_yolo26s.pt \
    --aed-root /path/to/AED --split test

Primary — point-based

A detection counts as correct when its centre lies within the tolerance, in metres, of an annotated elephant. 4.5 m is one nadir-visible body length, so 2.25 m is half a body length — the headline figure.

Tolerance	AP	Max recall	Best F1	@ conf	Precision	Recall
1.00 m (quarter body)	0.8445	0.8892	0.8563	0.50	0.8825	0.8316
2.25 m (half body)	0.9002	0.9276	0.8896	0.50	0.9168	0.8640
4.50 m (one body)	0.9094	0.9343	0.8930	0.50	0.9203	0.8673

At the 2.25 m operating point: 2,566 TP · 233 FP · 404 FN out of 2,970.

The metric is stable across tolerance — 0.845 → 0.909 for a 4.5× loosening — which says detections land close to the animal rather than merely nearby.

Postprocess sensitivity

Same detections, different cross-tile merge rule. SAHI defaults to the second row; use the first.

Postprocess	Point AP @ 2.25 m	Max recall	Best F1	@ conf	Precision	Recall
NMS / IOU (recommended)	0.9002	0.9276	0.8896	0.50	0.9168	0.8640
GREEDYNMM / IOS (SAHI default)	0.8532	0.8761	0.8752	0.45	0.9202	0.8343

IOS is intersection-over-smaller, so a small box inside a larger one always scores 1.0. Because this model emits altitude-scaled boxes, that over-merges distinct animals standing close together in herds — costing 0.047 AP and 0.05 recall.

Secondary — box mAP, stated with its convention

Postprocess	GT convention	mAP@0.5	mAP@0.5:0.95
NMS / IOU	4.5 m / GSD	0.8584	0.4651
NMS / IOU	3.5 m / GSD	0.6414	0.1589
GREEDYNMM / IOS	4.5 m / GSD	0.7381	0.3674
GREEDYNMM / IOS	3.5 m / GSD	0.4311	0.0989

Quote mAP@0.5, not mAP@0.5:0.95 — with synthesised ground truth the tight-IoU thresholds are scoring the size assumption, not the detector.

Baseline comparison

A YOLOv11s trained on a fixed 60 px box convention, run through the identical pipeline on the identical images (yolo11s_baseline.pt, included here):

	This model	Fixed-box baseline
Point AP @ 2.25 m, NMS/IOU	0.9002	0.8962
Point AP @ 2.25 m, GREEDYNMM/IOS	0.8532	0.8724
Best F1	0.8896 @ 0.50	0.8854 @ 0.45
Box mAP@0.5 (GT 4.5 m)	0.8584	0.1221
Predicted box width p10/median/p90	36 / 89 / 118 px	54 / 57 / 63 px
Required width p10/median/p90	37 / 80 / 113 px	37 / 80 / 113 px

These two models are statistically indistinguishable at finding elephants — and the ranking flips with the postprocess setting, so no detection win should be claimed for either. The real, large difference is localisation geometry: the baseline emits one box size across a 10.5× GSD range, this model tracks altitude.

Validation during training (tiles)

Precision 0.8896 / Recall 0.8472 / mAP50 0.8883 / mAP50-95 0.5304 at epoch 99 of 100. The validation split is 4,026 tiles from 325 held-out source photos in AED's training set, disjoint at photo level. All validation losses were still falling at epoch 100 and patience 25 never triggered — the run was not overfitting and had not converged.

Graphs


PR curve	Confidence sweep
Box size tracking	Box mAP conventions

box_size_tracking.png is the most diagnostic one: each dot is a test image, plotting the box width the model predicted against the width the altitude calls for. This model tracks the diagonal across the whole range.

Training


Base	`yolo26s.pt` (full 902/902 layer COCO transfer)
Epochs	100 (no early stop; patience 25 never triggered)
Image size	1024
Batch / workers	8 / 2 (12 GB RTX 3060)
Optimiser	auto, `lr0=0.01`, `lrf=0.01`, cosine, 5 warm-up epochs
Wall clock	22.2 h
Train set	17,385 tiles (5,795 positive + 11,590 background) from 1,304 photos
Val set	4,026 tiles from 325 photos

Labels are squares synthesised from AED points at 4.5 m / GSD — the nadir-visible longest dimension. (3.5 m, used by earlier work on this dataset, is an elephant's shoulder height, which is not what a top-down camera sees.)

degrees=0.0 is deliberate. The boxes are fixed-size squares. Rotation augmentation makes Ultralytics recompute the axis-aligned box of the rotated box, inflating it by up to √2 and corrupting the size convention. flipud=0.5 is enabled — from directly overhead, "up" is arbitrary.

Splits are disjoint at source-photo level, decided before tiling. Tiles overlap by 20%, so splitting tiles instead would put near-duplicate crops in both train and validation.

Files

File
`elefind_yolo26s.pt`	the model — use this
`yolo11s_baseline.pt`	fixed-box YOLOv11s, so the comparison above is reproducible
`assets/`	evaluation graphs
`metrics/corrected_metrics.{json,csv}`	every number above, machine-readable

Limitations and intended use

Intended for aerial wildlife survey research and conservation monitoring on nadir imagery of comparable resolution and terrain.

Known limitations:

One class. It detects elephants. It has never seen another species and will not distinguish one; expect false positives on large-bodied fauna.
Altitude range. Trained on GSD 0.024–0.130 m/px (roughly 320–1,712 m AGL). Outside that, behaviour is untested.
Domain shift is baked into the benchmark. AED's test sorties flew a median 745 m against 1,084 m for training, so test elephants are ~1.45× larger in pixels than anything in training. The reported numbers already include that penalty.
Boxes are synthetic. Predicted box extents approximate 4.5 m / GSD. Do not treat them as measurements of an animal.
Habitat. AED is savanna/bushland. Forest canopy, snow, and dense wetland are out of distribution.
Not for enforcement or targeting. This is a survey and monitoring tool. Counts carry real error (recall 0.864 at best F1) and should not be treated as a census without human verification.

Licence

AGPL-3.0. This is inherited, not chosen lightly: the weights derive from Ultralytics pretrained models and the inference path imports Ultralytics, which is AGPL-3.0. If you distribute a modified version or run one as a network service, you must publish your source.

The Aerial Elephant Dataset is not redistributed here. Obtain it from Zenodo under its own terms.

Citation

Training and evaluation code: https://github.com/iamhelitha/elefind

Please also cite the dataset:

bibtex

@inproceedings{naude2019aerial,
  title     = {The Aerial Elephant Dataset: A New Public Benchmark
               for Aerial Object Detection},
  author    = {Naud{\'e}, Johannes J. and Joubert, Deon},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern
               Recognition Workshops (CVPRW)},
  year      = {2019}
}