Object detection training

Training an Object Detection Model¶

When you're ready to train a model, the process generally boils down to three main steps:

Download the data for training, validation, and testing.
Transform that data into the format the model needs.
Initiate the training.

The instructions below show you how to train either RF-DETR(recommended) or the YOLOv26 model.

You will need the mbari-aidata package installed for this.

Tip

Grab your project's YAML file from the project page. If you haven't yet, install the aidata tool - see the installation instructions if you need help.

Setting Up¶

Download¶

Download the training data (object detection only). This excludes the held-out test set (testset) and optionally mixes in a percentage of unverified/negative background crops later during transform:

aidata download dataset \
  --config https://docs.mbari.org/internal/ai/projects/config/config_uav.yml \
  --base-path $PWD/TrainNegatives07152026 \
  --voc \
  --labels "Batray","Bird","Boat","Cement_Ship","Egregia","Fish","Jelly","Kayak","Kelp","Mola","Mooring_Buoy","Otter","Person","Pinniped","Secci_Disc","Shark","Surfboard","Velella_velella","Velella_velella_raft","Whale" \
  --verified \
  --token $TATOR_TOKEN \
  --disable-ssl-verify \
  --exclude-versions testset --single-class object

Download the testing data (object detection only), pulling only from the testset version:

aidata download dataset \
  --config https://docs.mbari.org/internal/ai/projects/config/config_uav.yml \
  --base-path $PWD/Test07152026 \
  --voc \
  --labels "Batray","Bird","Boat","Cement_Ship","Egregia","Fish","Jelly","Kayak","Kelp","Mola","Mooring_Buoy","Otter","Person","Pinniped","Secci_Disc","Shark","Surfboard","Velella_velella","Velella_velella_raft","Whale" \
  --verified \
  --token $TATOR_TOKEN \
  --disable-ssl-verify \
  --version testset

For more details, see the aidata download documentation.

Transform (train set)¶

Transform sizing depends on which model you're targeting:

YOLO models (all sizes) — use 640x640 images:

aidata transform voc --base-path $PWD/TrainNegatives07152026 --resize 640 --crop-size 640 --crop-overlap 0.5 --min-visibility 0.0 --min-dim 20 --negative-percent 0.2

RF-DETR Medium — use 576x576 crops. For other RF-DETR sizes, see the RF-DETR GitHub repo:

aidata transform voc --base-path $PWD/TrainNegatives07152026 --resize 576 --crop-size 576 --crop-overlap 0.5 --min-visibility 0.0 --min-dim 20

You can include background/negative crops with --negative-percent. A lower recommended value for RF-DETR:

aidata transform voc --base-path $PWD/TrainNegatives07152026 --resize 576 --crop-size 576 --crop-overlap 0.5 --min-visibility 0.0 --min-dim 20 --negative-percent 0.05

Convert to YOLO format:

aidata transform voc-to-yolo --base-path $PWD/TrainNegatives07152026/transformed

Split into train/validate/test sets. Since the held-out test set is downloaded separately above, the split here typically skips the test partition (--split 0.9,0.1,0.0):

aidata transform split -i $PWD/TrainNegatives07152026/transformed -o $PWD/TrainNegatives07152026split --split 0.9,0.1,0.0

For more details, see the transform command docs.

Compress the test data¶

Compress the held-out test set for easy upload into Google Drive (used later in Colab):

tar -czf images.tar.gz -C $PWD/Test07152026/testset images
tar -czf labels.tar.gz -C $PWD/Test07152026/testset labels

Train in AWS SageMaker with deepsea-ai 🚀¶

Training with deepsea-ai on AWS SageMaker has been moved to its own doc — see Training with deepsea-ai.

Training an RF-DETR Model in Google Colab Environment¶

Create a data.yaml file:

train: /content/datasets/train/images
val: /content/datasets/valid/images
test: /content/datasets/test/images

nc: 1
names: ['object']

Upload the data.yaml file, and the images.tar.gz/labels.tar.gz files (produced in the Compress the test data step, or from the train split) to Google Drive. Google Drive for Desktop makes this easy.

In a Colab notebook, install RF-DETR and dependencies:

%pip install "rfdetr[train,loggers]==1.7.1" -q
%pip install supervision roboflow -q

Set up HOME:

import os
HOME = os.getcwd()
print(HOME)

Mount Google Drive:

# Allow access to personal google drive and add new folders

# Connect Google Drive
from google.colab import drive
drive.mount("/content/drive", force_remount=True) # This will prompt for authorization.

# This will create the uavs files if they don't exist.
folders = ["uavs/"]
for folder in folders:
    path = "/content/drive/MyDrive/" + folder
    if not os.path.exists(path):  # Create the folder if it does not exist
        os.mkdir(path)

Move in the compressed files and unpack them:

!mkdir {HOME}/datasets
%cd {HOME}/datasets
from google.colab import userdata
uavs_folder = "/content/drive/MyDrive/uavs/"

!mkdir -p /content/datasets/savedir/
!cp -r "/content/drive/MyDrive/uavs/images.tar.gz" "/content/datasets/savedir/"
!cp -r "/content/drive/MyDrive/uavs/labels.tar.gz" "/content/datasets/savedir/"

!tar xf /content/datasets/savedir/images.tar.gz --directory /content/datasets/savedir/
!tar xf /content/datasets/savedir/labels.tar.gz --directory /content/datasets/savedir/

Move the data to the directory structure RF-DETR expects (note: RF-DETR uses valid/, not val/):

## create directories
!mkdir /content/datasets/train/
!mkdir /content/datasets/train/images/
!mkdir /content/datasets/train/labels/
!mkdir /content/datasets/test/
!mkdir /content/datasets/test/images/
!mkdir /content/datasets/test/labels/
!mkdir /content/datasets/valid/
!mkdir /content/datasets/valid/images/
!mkdir /content/datasets/valid/labels/

#get the data.yaml file
!cp "/content/drive/MyDrive/uavs/data.yaml" "/content/datasets/data.yaml"
!ls /content/datasets/

#move the data to the expected directories
!cp -r "/content/datasets/savedir/images/train/." "/content/datasets/train/images/"
!cp -r "/content/datasets/savedir/labels/train/." "/content/datasets/train/labels/"

!cp -r "/content/datasets/savedir/images/test/." "/content/datasets/test/images/"
!cp -r "/content/datasets/savedir/labels/test/." "/content/datasets/test/labels/"

!cp -r "/content/datasets/savedir/images/val/." "/content/datasets/valid/images/"
!cp -r "/content/datasets/savedir/labels/val/." "/content/datasets/valid/labels/"

!ls /content/datasets/

Train a new RF-DETR Nano model:

from rfdetr import RFDETRNano

model = RFDETRNano()

model.train(
    dataset_dir="/content/datasets/",
    epochs=20,
    batch_size=16,
    grad_accum_steps=1,
    output_dir="/content/outputs"
)

Copy outputs to Drive:

!cp -r "/content/outputs" "/content/drive/MyDrive/uavs/"

If Colab restarts, restore from Drive first:

!cp -r /content/drive/MyDrive/uavs/outputs /content/outputs

Resuming training from a checkpoint:

from rfdetr import RFDETRNano

model = RFDETRNano()
model.train(
    dataset_dir="/content/datasets/",
    epochs=50,           # total epochs — not "50 more", but the new total
    batch_size=16,
    grad_accum_steps=1,
    output_dir="/content/outputs",
    resume="/content/outputs/checkpoint_best_total.pth"   # <-- resume here
)

Exporting the model to ONNX format:

model = RFDETRNano(pretrain_weights="/content/outputs/checkpoint_best_total.pth")
model.export(
    format="onnx",
    output_dir="/content/drive/MyDrive/uavs/onnx_export"
)

Training a YOLO26 Model in Google Colab Environment¶

Create the data.yaml file:

train: /content/datasets/train/images
val: /content/datasets/val/images
test: /content/datasets/test/images

nc: 1
names: ['object']

Upload the data.yaml file, and the image/label tar files to Google Drive. Google Drive for Desktop makes this easy.

Install YOLO26 via Ultralytics:

%pip install ultralytics supervision roboflow -q
import ultralytics
ultralytics.checks()

Set up HOME:

import os
HOME = os.getcwd()
print(HOME)

Mount Google Drive:

# Allow access to personal google drive and add new folders

# Connect Google Drive
from google.colab import drive
drive.mount("/content/drive", force_remount=True) # This will prompt for authorization.

# This will create the uavs files if they don't exist.
folders = ["uavs-yolo26/"]
for folder in folders:
    path = "/content/drive/MyDrive/" + folder
    if not os.path.exists(path):  # Create the folder if it does not exist
        os.mkdir(path)

Move in the compressed files and unpack them:

!mkdir {HOME}/datasets
%cd {HOME}/datasets
from google.colab import userdata
uavs_folder = "/content/drive/MyDrive/uavs-yolo26/"

!mkdir -p /content/datasets/savedir/
!cp -r "/content/drive/MyDrive/uavs-yolo26/images.tar.gz" "/content/datasets/savedir/"
!cp -r "/content/drive/MyDrive/uavs-yolo26/labels.tar.gz" "/content/datasets/savedir/"

!tar xf /content/datasets/savedir/images.tar.gz --directory /content/datasets/savedir/
!tar xf /content/datasets/savedir/labels.tar.gz --directory /content/datasets/savedir/

Move the data to the directory structure YOLO26 expects:

## make the directories that yolo26 expects
!mkdir /content/datasets/train/
!mkdir /content/datasets/train/images/
!mkdir /content/datasets/train/labels/
!mkdir /content/datasets/test/
!mkdir /content/datasets/test/images/
!mkdir /content/datasets/test/labels/
!mkdir /content/datasets/val/
!mkdir /content/datasets/val/images/
!mkdir /content/datasets/val/labels/

#get the data.yaml file
!cp "/content/drive/MyDrive/uavs-yolo26/data.yaml" "/content/datasets/data.yaml"
!ls /content/datasets/

#move the data to the expected directories
!cp -r "/content/datasets/savedir/images/train/" "/content/datasets/train/images/"
!cp -r "/content/datasets/savedir/labels/train/" "/content/datasets/train/labels/"

!cp -r "/content/datasets/savedir/images/test/" "/content/datasets/test/images/"
!cp -r "/content/datasets/savedir/labels/test/" "/content/datasets/test/labels/"

!cp -r "/content/datasets/savedir/images/val/" "/content/datasets/val/images/"
!cp -r "/content/datasets/savedir/labels/val/" "/content/datasets/val/labels/"

!ls /content/datasets/

Train a YOLO26 Medium model for 150 epochs:

!yolo task=detect mode=train model=yolo26m.pt data=/content/datasets/data.yaml \
  batch=128 epochs=150 patience=30 imgsz=640 \
  mixup=0.3 scale=0.9 plots=True

Save outputs to Drive:

!cp "/content/datasets/runs/detect/train/weights/best.pt" "/content/drive/MyDrive/uavs-yolo26/best.pt"
!cp "/content/datasets/runs/detect/train/weights/last.pt" "/content/drive/MyDrive/uavs-yolo26/last.pt"
!cp -r "/content/datasets/runs/detect/train/" "/content/drive/MyDrive/uavs-yolo26/train/"

You can train the model again for more epochs by loading the last checkpoint. Copy the last checkpoint from Drive:

!cp /content/drive/MyDrive/uavs-yolo26/last.pt /content/last.pt

Train for another 150 epochs starting from the last checkpoint:

!yolo task=detect mode=train model=/content/last.pt data=/content/datasets/data.yaml \
  batch=128 epochs=150 patience=30 imgsz=640 \
  mixup=0.3 scale=0.9 plots=True

Export the model to ONNX format:

from ultralytics import YOLO

model = YOLO("/content/drive/MyDrive/uavs-yolo26/best.pt")
model.export(format="onnx")

🗓️ Updated: 2026-08-10