🚀 Object Instance Segmentation Using YOLOv11s and a 10 Image Dataset

🌟 Project Overview

This project focuses on building and training an advanced model for Instance Segmentation using the YOLOv11s architecture. This specific model was chosen to achieve the best balance between speed, lightweightness, and accuracy compared to other versions.

This project focuses on building and training an advanced Instance Segmentation model using the YOLOv11s architecture, which was selected to achieve the best balance between speed, lightweightness, and accuracy compared to other versions. The project is based on a small dataset consisting of only 10 images.

• The YOLOv11s model is distinguished by being highly compact, ensuring it runs smoothly even on devices with limited resources.
• The low Inference Time makes it ideal for Real-Time Applications.
• The training process was straightforward and required minimal memory on a T4 GPU.

---

⚙️ Model Setup and Configuration

Parameter	Value	Description	Source
Task	segment	Specifies the instance segmentation task
Model	yolo11s-seg.pt	The selected model (YOLOv11s)
Epochs	150	Maximum number of training epochs
Patience	20	Number of epochs to wait for improvement before Early Stopping
Batch Size	16	Number of images in each training step
imgsz	640	Image resolution used for training and inference

---

🧪 Data Pipeline & Validation Strategy

The project relied on a comprehensive data processing pipeline to compensate for the scarcity of the original dataset.

1. 🖼️ Tiling and Conversion Pipeline

A pipeline was implemented to convert COCO data to YOLO format, focusing on the Tiling technique to:

• Effectively increase the number of training images without collecting new data.
• Enable the model to see smaller details more clearly.
• Use SAHI Slicing to divide COCO images and correctly adjust their Annotations into tiles of size 640×640.
• Conversion to the YOLO format was used with the use_segments=True option to create .txt files and correctly save the Segmentation Masks.

2. 🎨 Data Augmentation

Since the number of images after tiling was small (about 300 images), data augmentation was used to increase variety and scene coverage:

• Tool: The Roboflow platform was relied upon for its ready-made and organized tools, which simplified the process.
• Applied Transformations: Rotation, resizing, brightness adjustment, application of Flips, and combining multiple transformations.

3. 📊 Data Splitting Strategy

• Ratio: The dataset was split with a ratio of 0.8 for Training and 0.2 for Validation.
• Test Set: A separate test set was not allocated due to data scarcity, and the Validation Set was relied upon to evaluate the model's final performance.
• Quality: Care was taken to prevent Data Leakage to ensure reliable evaluation, as the validation images were never seen by the model.

---

📈 Key Results and Performance

Strong and reliable results were obtained, as the Loss Curves were smooth, and validation metrics remained close to training metrics, indicating reliable model training without Overfitting.

The following table summarizes the best Mask Segmentation performance on the Validation Set:

Metric (Mask Segmentation)	Value (Best/Final)	Significance
mAP@0.5	0.98533 (Final Epoch 117)	High average precision exceeding 98%.
Precision	0.99605 (Final Epoch 117)	Near-perfect precision (approaching 1.00).
Recall	0.96338 (Best Epoch 111)	High recall (around 95.8%) indicating detection of the majority of objects.
mAP@0.5:0.95	0.90066 (Best Epoch 111)	Excellent average precision across a wide range of IoU thresholds.

Conclusion: These results prove that YOLOv11s is a practical and trustworthy choice for this task.