Sea Turtles Detection

Detect sea turtles from drone imagery using YOLO object detection models.

Overview

This project trains and benchmarks multiple YOLO architectures for detecting sea turtles in aerial/drone images. It uses a dataset from Roboflow containing annotated NIR (near-infrared) drone footage of sea turtles on nesting beaches.

Dataset

Two dataset variants were used across experiments:

ID	Name	Train	Val	Test	Notes
1	Base (B)	968	272	118	Original annotations
2	Tiled (W)	15,488	4,352	1,888	Tiled augmentation

Class: Turtle (single-class detection)
Source: Roboflow — sea-turtles-yia2e
License: CC BY 4.0

The dataset is stored locally under Dataset/sea-turtles-1/ and is excluded from version control (see .gitignore).

Benchmark Results

All models trained for 100 epochs on Dataset 1 (Base), image size 640, batch 8, seed 42.

Model	mAP50	mAP50-95	Precision	Recall	Train Time (min)
yolov9c	0.955	0.633	0.935	0.910	106.8
yolo26m	0.952	0.643	0.927	0.897	93.2
yolov9m	0.950	0.619	0.913	0.900	95.4
yolov11m	0.946	0.616	0.896	0.915	81.0
yolov8m	0.946	0.613	0.922	0.910	78.2
yolov10m	0.944	0.619	0.924	0.903	90.8
yolo26s	0.940	0.593	0.860	0.913	69.6
yolov10s	0.936	0.589	0.892	0.893	67.6
yolov8s	0.936	0.567	0.905	0.866	54.7
yolov11s	0.932	0.574	0.892	0.882	56.5
yolov9s	0.925	0.559	0.875	0.856	89.9
yolov11n	0.907	0.524	0.857	0.838	54.2
yolov10n	0.907	0.525	0.845	0.854	63.1
yolo26n	0.887	0.508	0.832	0.831	70.8
yolov9t	0.886	0.497	0.840	0.808	86.8
yolov8n	0.886	0.500	0.854	0.794	50.6

Full results: yolo_ultralytics_benchmark/results/benchmark_results.csv

Earlier Experiments

Run	Architecture	Epochs	Precision	Recall	mAP50	Dataset	Notes
1	YOLOv8n	50	0.748	0.764	0.809	1	Colab
2	YOLOv8n	5	0.652	0.419	0.453	2	Colab
3	YOLOv8n	150	0.830	0.920	0.825	1	Local
4	YOLOv8n	400	0.824	0.920	0.823	1	Local
5	YOLOv8n	100	0.786	0.910	0.827	2	Local
7	YOLOv8m	100	0.866	0.666	0.796	1	Local

Project Structure

sea-turtles-detection/
├── Dataset/
│   └── sea-turtles-1/          # YOLO-format dataset (gitignored)
│       ├── train/images/
│       ├── valid/images/
│       ├── test/images/
│       └── data.yaml
│
├── yolo_ultralytics_benchmark/ # Multi-model benchmark (v8/v9/v10/v11/v26)
│   ├── models/                 # Pre-trained .pt weights (gitignored)
│   ├── scripts/
│   │   ├── benchmark_yolo_models.py  # Main benchmark runner
│   │   └── check_models.py
│   ├── results/
│   │   └── benchmark_results.csv
│   └── runs/                   # Training outputs (gitignored)
│
├── archive/                    # Previous experiments
│   ├── yolov8/                 # Colab notebooks (train/val/detect)
│   ├── yolov8_local/           # Local notebooks + hyperparameter tuning
│   ├── yolov8_tile/            # Tiled dataset experiments (Colab)
│   ├── yolov8_tile_local/      # Tiled dataset experiments (local)
│   ├── yolov5_local/           # YOLOv5 baseline
│   ├── yolov9/                 # YOLOv9 training notebooks
│   ├── yolov9_ultralytics/     # YOLOv9 via Ultralytics API
│   └── yolov26n_ultralytics/   # YOLO26n via Ultralytics API
│
├── runs/                       # Detection inference outputs (gitignored)
├── env.sample                  # Environment variable template
└── Makefile                    # venv activation helpers

Setup

Prerequisites

Python 3.8+
CUDA-capable GPU (recommended)

Install dependencies

python -m venv .venv
source .venv/Scripts/activate   # Windows
# source .venv/bin/activate     # Linux/macOS
pip install ultralytics

Environment variables

Copy env.sample to .env and fill in your API keys:

cp env.sample .env

Required keys:

ROBOFLOW_API_KEY — for dataset access
COMET_API_KEY — for experiment tracking (optional)

Running the Benchmark

cd yolo_ultralytics_benchmark
python scripts/benchmark_yolo_models.py

Results are appended to yolo_ultralytics_benchmark/results/benchmark_results.csv.

Pre-trained model weights must be placed in yolo_ultralytics_benchmark/models/ before running.

Running Inference

yolo detect predict model=yolo_ultralytics_benchmark/models/yolov9c.pt source=<image_or_folder>

Training Results Analysis

All 16 models were trained for 100 epochs on the same dataset and hardware. Results are stored in yolo_ultralytics_benchmark/results/images/.

Summary Table

Model	mAP50	mAP50-95	Precision	Recall	Train Time (min)	Efficiency (mAP50/min)
yolov9c	0.955	0.633	0.935	0.910	106.8	0.00894
yolo26m	0.952	0.643	0.927	0.897	93.2	0.01022
yolov9m	0.950	0.619	0.913	0.900	95.4	0.00996
yolov11m	0.946	0.616	0.896	0.915	81.0	0.01168
yolov8m	0.946	0.613	0.922	0.910	78.2	0.01210
yolov10m	0.944	0.619	0.924	0.903	90.8	0.01040
yolo26s	0.940	0.593	0.860	0.913	69.6	0.01351
yolov10s	0.936	0.589	0.892	0.893	67.6	0.01385
yolov8s	0.936	0.567	0.905	0.866	54.7	0.01711
yolov11s	0.932	0.574	0.892	0.882	56.5	0.01650
yolov9s	0.925	0.559	0.875	0.856	89.9	0.01029
yolov11n	0.907	0.524	0.857	0.838	54.2	0.01674
yolov10n	0.907	0.525	0.845	0.854	63.1	0.01437
yolo26n	0.887	0.508	0.832	0.831	70.8	0.01253
yolov9t	0.886	0.497	0.840	0.808	86.8	0.01021
yolov8n	0.886	0.500	0.854	0.794	50.6	0.01751

Efficiency = mAP50 / train time. Higher is better for resource-constrained deployments.

Key Observations

Best accuracy: YOLOv9c achieves the highest mAP50 (0.955) and strong precision (0.935), at the cost of the longest training time (106.8 min).
Best mAP50-95: YOLO26m leads at 0.643, meaning it localises turtles more precisely (tighter bounding boxes).
Best efficiency: YOLOv8n delivers the fastest training (50.6 min) and competitive mAP50 (0.886) — best for rapid iteration.
Best small model: YOLOv8s reaches 0.936 mAP50 in only 54.7 min, the most efficient small model overall.
YOLOv9 family: Consistently strong across all sizes; the gap between t → s → m → c is large (~0.040 mAP50), making model size selection important.
YOLO11 vs YOLOv8: YOLO11 trains in similar time but achieves higher recall across all sizes, suggesting better detection coverage with fewer misses.
YOLO26: The newest architecture; the medium variant rivals YOLOv9c in mAP50-95 while training ~13 min faster.
Nano models (v8n, v9t, v10n, v11n, v26n): All cluster around 0.886–0.907 mAP50. YOLOv11n and YOLOv10n lead the nano tier.

Per-Model Results

YOLOv9c — Best Overall (mAP50: 0.955)

Metric	Value
mAP50	0.955
mAP50-95	0.633
Precision	0.935
Recall	0.910
Train time	106.8 min

Training curves show steady convergence with no overfitting. Loss continues decreasing through epoch 100.

YOLO26m — Best mAP50-95 (0.643)

Metric	Value
mAP50	0.952
mAP50-95	0.643
Precision	0.927
Recall	0.897
Train time	93.2 min

Highest mAP50-95 of all models, indicating the tightest bounding box localisation. Trains ~13 min faster than YOLOv9c for comparable mAP50.

YOLOv9m

Metric	Value
mAP50	0.950
mAP50-95	0.619
Precision	0.913
Recall	0.900
Train time	95.4 min

Very close to YOLOv9c in accuracy (~0.005 mAP50 gap) but with a smaller model. A strong choice when model size matters.

YOLO11m

Metric	Value
mAP50	0.946
mAP50-95	0.616
Precision	0.896
Recall	0.915
Train time	81.0 min

Best recall among medium models (0.915), meaning fewest missed turtles. Trains significantly faster than YOLOv9c/m for only ~0.009 mAP50 loss.

YOLOv8m

Metric	Value
mAP50	0.946
mAP50-95	0.613
Precision	0.922
Recall	0.910
Train time	78.2 min

Fastest of the medium models. Matches YOLO11m in mAP50 while training ~3 min faster, with higher precision but slightly lower recall.

YOLOv10m

Metric	Value
mAP50	0.944
mAP50-95	0.619
Precision	0.924
Recall	0.903
Train time	90.8 min

Highest precision among medium models (0.924). Good mAP50-95 matching YOLOv9m, but takes longer to train than v8m/v11m for similar mAP50.

YOLO26s

Metric	Value
mAP50	0.940
mAP50-95	0.593
Precision	0.860
Recall	0.913
Train time	69.6 min

Impressive for a small model — nearly matches YOLOv10m mAP50 in 69.6 min. High recall (0.913) makes it a strong candidate when minimising missed detections.

YOLOv10s

Metric	Value
mAP50	0.936
mAP50-95	0.589
Precision	0.892
Recall	0.893
Train time	67.6 min

Balanced small model. Well-rounded precision/recall with fast training.

YOLOv8s — Best Small Model Efficiency

Metric	Value
mAP50	0.936
mAP50-95	0.567
Precision	0.905
Recall	0.866
Train time	54.7 min

Highest mAP50 per minute among small models. Excellent choice when training time and accuracy must both be minimised.

YOLO11s

Metric	Value
mAP50	0.932
mAP50-95	0.574
Precision	0.892
Recall	0.882
Train time	56.5 min

Slightly below YOLOv8s in mAP50 but better mAP50-95 (0.574 vs 0.567), indicating marginally better localisation.

YOLOv9s

Metric	Value
mAP50	0.925
mAP50-95	0.559
Precision	0.875
Recall	0.856
Train time	89.9 min

Weakest small model — takes nearly as long as the medium YOLOv9m (~95 min) but achieves considerably lower accuracy. YOLOv8s or YOLO11s are better choices at this tier.

YOLO11n

Metric	Value
mAP50	0.907
mAP50-95	0.524
Precision	0.857
Recall	0.838
Train time	54.2 min

Best nano model alongside YOLOv10n. Fastest training of all models.

YOLOv10n

Metric	Value
mAP50	0.907
mAP50-95	0.525
Precision	0.845
Recall	0.854
Train time	63.1 min

Ties YOLO11n in mAP50 but trains ~9 min slower. Marginally better mAP50-95.

YOLO26n

Metric	Value
mAP50	0.887
mAP50-95	0.508
Precision	0.832
Recall	0.831
Train time	70.8 min

Weakest YOLO26 variant. Slower than YOLO11n while delivering lower accuracy — the larger YOLO26s/m are far more competitive.

YOLOv9t

Metric	Value
mAP50	0.886
mAP50-95	0.497
Precision	0.840
Recall	0.808
Train time	86.8 min

Slowest and least accurate nano model. The tiny variant of YOLOv9 does not benefit from training time vs accuracy trade-off on this dataset.

YOLOv8n — Fastest Training

Metric	Value
mAP50	0.886
mAP50-95	0.500
Precision	0.854
Recall	0.794
Train time	50.6 min

Fastest training time overall. Low recall (0.794) means it misses more turtles than other models — not ideal for conservation monitoring where false negatives carry high cost.

Recommendations

Use case	Recommended model
Best accuracy	YOLOv9c
Best accuracy + efficiency	YOLO26m or YOLOv8m
Resource-constrained deployment	YOLOv8s
Fastest iteration / prototyping	YOLO11n or YOLOv8n
Minimise missed detections (high recall)	YOLO11m or YOLO26s

Editing Colab Notebooks

Create a new branch on GitHub.
Open the notebook via: https://colab.research.google.com/github/gab-es21/sea-turtles-detection/ (select your branch)
File > Save a copy in Drive, make edits, then File > Save a copy in GitHub.

License

Dataset licensed under CC BY 4.0. See LICENSE for repository license.

Name		Name	Last commit message	Last commit date
Latest commit History 94 Commits
Dataset/sea-turtles-2		Dataset/sea-turtles-2
archive		archive
configs		configs
docs		docs
results		results
scripts		scripts
.gitignore		.gitignore
ANALYSIS.md		ANALYSIS.md
Dataset-20231029T233312Z-001.zip		Dataset-20231029T233312Z-001.zip
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
ROADMAP.md		ROADMAP.md
env.sample		env.sample
tree.txt		tree.txt

Folders and files

Latest commit

History

Repository files navigation

Sea Turtles Detection

Overview

Dataset

Benchmark Results

Earlier Experiments

Project Structure

Setup

Prerequisites

Install dependencies

Environment variables

Running the Benchmark

Running Inference

Training Results Analysis

Summary Table

Key Observations

Per-Model Results

YOLOv9c — Best Overall (mAP50: 0.955)

YOLO26m — Best mAP50-95 (0.643)

YOLOv9m

YOLO11m

YOLOv8m

YOLOv10m

YOLO26s

YOLOv10s

YOLOv8s — Best Small Model Efficiency

YOLO11s

YOLOv9s

YOLO11n

YOLOv10n

YOLO26n

YOLOv9t

YOLOv8n — Fastest Training

Recommendations

Editing Colab Notebooks

License

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

Packages