The model was trained using a reproducible MLflow-based experimentation pipeline, and the best-performing checkpoint has been released for inference and downstream use.

🔍 Model Summary

Task: Instance Segmentation
Architecture: YOLO (Ultralytics)
Released Model: YOLO26n-seg
Input: RGB images
Output: Bounding boxes + segmentation masks
Framework: Ultralytics YOLO

📊 Dataset

Source: ROBOFLOW Bottle Detection
Classes: 1
- ['bottle-JBay']
Dataset splits:
- Train: 587 images, 845 instances
- Valid: 167 images, 239 instances
- Test: 84 images, 119 instances

📈 Evaluation Results

Validation Set

Metric	Value
Mask mAP50	0.952
Mask mAP50–95	0.808
Box mAP50	0.952
Box map50-95	0.795

Test Set

Metric	Value
Mask mAP50	0.963
Mask mAP50–95	0.864
Box mAP50	0.963
Box map50-95	0.807

Detailed metrics are available in:

results_valid.csv
results_test.csv

🧪 Training Details

Image size: 640 × 640
Optimizer: Adam
Batch size: 8
Early stopping: Enabled
Hardware: Single NVIDIA GPU (8 GB)
Experiment tracking: MLflow

The full training configuration is provided in config.yaml.

🚀 Usage

Installation

pip install ultralytics

Sample Inference

from ultralytics import YOLO

model = YOLO("yolo26_bottle_segmentation.pt")
results = model("image.jpg")

results[0].show()

Sample Output

Below is an example segmentation output produced by the model:

abhishekdey/bottle-yolo26-seg

🚀 Live Demo (Try it out)

Bottle Instance Segmentation – YOLO26