People Instance Segmentation using YOLOv8

A lightweight computer vision project that performs instance segmentation of people using Ultralytics YOLOv8. The repository demonstrates an end-to-end deep learning workflow—from dataset preparation and annotation conversion to model training, evaluation, and inference on both images and video.

---

Demo

---

Project Highlights

• Trained a YOLOv8n-Seg model for people instance segmentation.
• Converted annotations from COCO format to YOLO segmentation format.
• Performed dataset verification and train/validation/test splitting.
• Achieved high segmentation accuracy while remaining suitable for CPU inference.
• Demonstrated inference on both images and video.
• Organised into a clean, reproducible project structure.

---

Qualitative Results

---

Model Performance

Bounding Box Performance

Metric	Score
Precision	0.991
Recall	0.956
mAP@50	0.977
mAP@50-95	0.885

Segmentation Performance

Metric	Score
Precision	0.988
Recall	0.952
mAP@50	0.968
mAP@50-95	0.848

The trained model demonstrated strong localisation and segmentation performance on the validation dataset while remaining lightweight enough for CPU-based inference.

---

Repository Structure

people_instance_segmentation/
├── assets/
├── data/
├── dataset/
├── models/
├── outputs/
├── reports/
├── scripts/
├── videos/
├── data.yaml
├── requirements.txt
├── README.md
├── LICENSE
└── .gitignore

---

Installation

Clone the repository

git clone https://github.com/only1jamjam-ctrl/people_instance_segmentation-yolov8.git

Move into the project

cd people_instance_segmentation

Install the required packages

pip install -r requirements.txt

---

Training

yolo segment train \
model=yolov8n-seg.pt \
data=data.yaml \
epochs=50 \
imgsz=640 \
batch=8 \
project=runs \
name=people_seg

---

Inference

Images

yolo segment predict \
model=models/best.pt \
source=data/images/test \
save=True

Video

yolo segment predict \
model=models/best.pt \
source=outputs/demovideo.mp4 \
save=True

---

Project Workflow

flowchart LR

A[Dataset] --> B[COCO to YOLO Conversion]
B --> C[Dataset Verification]
C --> D[Train / Validation / Test Split]
D --> E[YOLOv8 Training]
E --> F[Model Evaluation]
F --> G[Image Prediction]
G --> H[Video Prediction]
H --> I[GIF Generation]

Loading

---

Key Achievements

• Trained a YOLOv8n-Seg model for 50 epochs.
• Achieved 0.968 Mask mAP@50 and 0.977 Box mAP@50.
• Built a complete COCO → YOLO segmentation preprocessing pipeline.
• Performed image and video inference using the trained model.
• Generated deployment-ready prediction outputs and demonstration GIF.
• Designed a clean, reproducible repository structure for future development.

---

Limitations

Although the model performs well on standard scenes, some challenging scenarios remain:

• Crowded scenes with heavy occlusion may reduce detection accuracy.
• Human-like objects such as dolls or mannequins can occasionally produce false positives.
• Performance can be further improved using a larger and more diverse training dataset.

---

Future Work

• Train using a larger dataset.
• Evaluate larger YOLOv8 segmentation backbones.
• Improve robustness on crowded scenes.
• Reduce false positives through hard-negative examples.
• Deploy the model for real-time webcam inference.

---

License

Released under the MIT License.