Update README.md

## 🤝 Thanks / Contact

For improvements, issues, or contributions — feel free to open a Pull Request or Discussion.

**Author:** **Tanish Jain**

README.md

@@ -1,21 +1,39 @@
 ---
 license: mit
+tags:
+- yolo
+- yolov8
+- object-detection
+- ppe
+- safety-detection
+- onnx
+- jetson
+- edge-ai
 ---
-# Model Card: YOLOv8n PPE Detection (6 Classes)
 
-## 1. Model Details
+# 🦺 YOLOv8n PPE Detection (6 Classes)
 
-- **Model name**: YOLOv8n PPE Detection – 6 Classes  
-- **Architecture**: Ultralytics YOLOv8n (nano variant)  
-- **Task**: Object Detection (PPE / safety gear)  
-- **Framework**: PyTorch (Ultralytics), exported to ONNX  
-- **Input size**: `1 × 3 × 640 × 640` (RGB, normalized 0–1)  
-- **Output shape (ONNX)**: `1 × 10 × 8400`  
-  - 4 channels: bounding box (x, y, w, h)  
-  - 6 channels: class probabilities  
-- **Number of classes**: 6  
+A lightweight and fast Personal Protective Equipment (PPE) detection model trained on 6 safety classes, optimized for edge devices such as **NVIDIA Jetson Orin**, **ONNX Runtime**, and **TensorRT** deployment.
 
-### Classes
+---
+
+# 1. 📌 Model Details
+
+- **Model name:** YOLOv8n PPE Detection – 6 Classes  
+- **Architecture:** Ultralytics YOLOv8n (Nano)  
+- **Task:** Object Detection (PPE / Safety Gear)  
+- **Framework:** PyTorch (Ultralytics), exported to ONNX  
+- **Input size:** `1 × 3 × 640 × 640`  
+- **Output shape (ONNX):** `1 × 10 × 8400`  
+  - 4 channels → bounding box (x, y, w, h)  
+  - 6 channels → class logits  
+- **Number of classes:** `6`  
+
+---
+
+# 2. 🏷️ Classes & Dataset Distribution
+
+### **Class List with Image Counts**
 
 | ID | Class Name   | Image Count |
 |----|--------------|-------------|
@@ -28,76 +46,86 @@ license: mit
 
 ---
 
-## 2. Intended Use
+# 3. 🎯 Intended Use
 
-### Primary use-cases
+### **Primary use-cases**
 
-- Workplace safety monitoring on:
-  - Construction sites
-  - Industrial plants
-  - Warehouses and logistics hubs
-- Real-time PPE compliance monitoring:
-  - Detect whether workers wear helmets, vests, gloves, masks, goggles, and safety shoes.
+- Real-time PPE compliance monitoring  
+- Construction site safety analytics  
+- Industrial / warehouse safety automation  
+- Worker equipment detection on:
+  - Helmets  
+  - Safety shoes  
+  - Reflective vests  
+  - Gloves  
+  - Masks  
+  - Goggles  
 
-### Target platforms
+### **Target Platforms**
 
-- NVIDIA Jetson Orin / Orin Nano / Xavier / Laptop
-- Edge devices using:
-  - ONNX Runtime
-  - TensorRT (FP16 / INT8)
-  - NVIDIA DeepStream
+- NVIDIA Jetson (Orin Nano, Orin NX, Xavier NX)
+- ONNX Runtime
+- TensorRT (FP32 / FP16 / INT8)
+- NVIDIA DeepStream SDK
+- Python (Ultralytics inference)
 
-### Not intended for
+### **Not intended for**
 
-- Legal or regulatory compliance **without human supervision**
-- Medical or clinical decisions
-- Surveillance in sensitive environments without proper consent/authorization
+- Legal enforcement without human review  
+- Medical or clinical decisions  
+- Surveillance without consent  
+- Environments restricted by privacy regulations  
 
-**Important**: This model should be used as a **support tool**. Final decisions should always be made by a human.
+> ⚠️ **Human supervision is mandatory for high-risk applications.**
 
 ---
 
-## 3. Training Data
+# 4. 📂 Training Data
 
-- **Dataset type**: PPE detection dataset (6 classes)  
-- **Images**: mix of indoor/outdoor scenes, workers wearing different PPE combinations  
-- **Annotation format**: YOLO format (class_id x_center y_center width height, normalized)  
-- **Train / Validation / Test split**:
-  - Train: majority of images  
-  - Validation: held-out subset for tuning  
-  - Test: separate hold-out, not used in training  
+- **Dataset source:** Custom PPE dataset (Roboflow format)  
+- **Annotation format:** YOLO TXT  
+- **Data types:** Indoor/outdoor industrial environments  
+- **Split:**
+  - Train: majority
+  - Val: held-out validation
+  - Test: separate set for unbiased evaluation  
 
-Data is skewed slightly toward **Vest** and **Gloves**, with fewer examples for **safety_shoe** and **goggles**, which can affect per-class performance.
+Dataset is slightly imbalanced:
+- **Vest** is most common  
+- **safety_shoe** & **goggles** have fewer samples  
 
 ---
 
-## 4. Training Procedure
-
-- **Base model**: `yolov8n.pt` (pretrained on COCO)  
-- **Training script**: Ultralytics `yolo train`  
-- **Key hyperparameters**:
-  - `epochs`: 50  
-  - `imgsz`: 640  
-  - `batch`: 32 (T4 GPU)  
-  - `optimizer`: auto (SGD/AdamW depending on Ultralytics config)  
-- **Loss & augmentations**:  
-  - Default Ultralytics YOLOv8 settings  
-  - Mosaic, random flip, color jitter (HSV), scaling, translation
+# 5. ⚙️ Training Procedure
+
+- **Base model:** `yolov8n.pt` (COCO pretrained)  
+- **Framework:** Ultralytics YOLOv8  
+- **Hyperparameters:**
+  - `epochs = 50`
+  - `imgsz = 640`
+  - `batch = 32`
+  - `optimizer = auto`
+- **Augmentations:**
+  - Mosaic
+  - HSV color shift
+  - Random flip
+  - Scaling
+  - Translation  
 
 ---
 
-## 5. Evaluation & Metrics
+# 6. 📊 Evaluation & Metrics
 
-Validation was run with Ultralytics default `val` pipeline.
+### **Overall Performance**
 
-**Overall performance:**
+| Metric          | Value |
+|-----------------|-------|
+| mAP@50          | ~0.81 |
+| mAP@50–95       | ~0.53 |
+| Precision       | ~0.80 |
+| Recall          | ~0.74 |
 
-- **mAP@50**: ~0.81  
-- **mAP@50–95**: ~0.53  
-- **Precision**: ~0.80  
-- **Recall**: ~0.74  
-
-**Per-class mAP@50 (approx)**:
+### **Per-Class mAP@50**
 
 | Class        | mAP@50 |
 |--------------|--------|
@@ -108,46 +136,37 @@ Validation was run with Ultralytics default `val` pipeline.
 | mask         | ~0.80  |
 | safety_shoe  | ~0.64  |
 
-These numbers are approximate and can vary slightly based on exact validation configuration.
-
 ---
 
-## 6. Limitations & Ethical Considerations
+# 7. ⚠️ Limitations & Ethical Considerations
 
-### Technical limitations
+### **Technical**
 
-- Performance may degrade in:
-  - Very low light  
-  - Motion blur (fast movement)  
-  - Heavy occlusion (crowded scenes)  
-  - Extremely small objects (far-away workers)
-- Model expects image shape 640×640; non-square / very different aspect ratios may slightly reduce accuracy.
+- Sensitive to low-light & motion blur  
+- Small distant objects can reduce accuracy  
+- PPE variations (colors, styles, regional differences) may affect performance
 
-### Data & bias
+### **Bias & Data Risks**
 
-- Training data may not represent **all industries, uniforms, skin tones, or environments** globally.  
-- PPE styles (helmet colors, vest types, shoe styles) may differ from the training distribution, impacting accuracy.
+- Dataset may not represent all ethnicities, industries, or regions globally  
+- Model may fail on unseen uniform types or PPE color variations  
 
-### Responsible use
+### **Responsible AI Guidance**
 
-- Do **not** use this model as the only mechanism to enforce safety or compliance.  
-- Always include:
-  - Human oversight  
-  - Clear communication to workers about monitoring  
-  - Respect for privacy and local regulations
+- Always include human review  
+- Follow local privacy laws (GDPR, workplace regulations)  
+- Use only for transparent, ethical monitoring  
 
 ---
 
-## 7. How to Use
+# 8. 🚀 How to Use
 
-### Ultralytics (.pt) – Python example
+## **A) Inference with Ultralytics (.pt)**
 
 ```python
 from ultralytics import YOLO
 
 model = YOLO("best.pt")
-results = model("image.jpg", imgsz=640, conf=0.25)
+results = model("image.jpg", imgsz=640)
 
-for r in results:
-    print(r.boxes.cls, r.boxes.conf, r.boxes.xyxy)
-    r.show()  # show window
+results[0].show()  # visualize