---
license: mit
language:
- en
metrics:
- accuracy
pipeline_tag: robotics
tags:
- robotics
- opencv
- tpu
- gpu
- tensorflow
---

# 🧠⚡ Protoge-LG: Scalable Object Detection & Tracking (GPU/TPU-Ready)

[![TensorFlow](https://img.shields.io/badge/framework-TensorFlow-orange)](https://www.tensorflow.org/)
[![Detection & Tracking](https://img.shields.io/badge/task-Detection%20%26%20Tracking-blue)]()
[![Labels](https://img.shields.io/badge/labels-1000%2B-green)]()
[![Accelerated](https://img.shields.io/badge/accelerator-GPU%2FTPU-red)]()

**Protoge-LG** is a large-scale object detection and tracking algorithm using computer vision and TensorFlow. Optimized for **GPU and TPU inference**, this model can detect and track over **1000 object classes** in real-time, with flexibility for targeted tracking of selected labels.

---

## 🚀 Highlights

- 🧠 Detects & tracks **1000+ object categories**
- 🌀 Supports both **full-label** and **targeted detection modes**
- ⚡ Accelerated with **GPU or TPU support**
- 🧰 Uses TensorFlow’s object detection API and integrates with OpenCV for real-time video processing
- 📦 Exportable to TensorFlow Lite, TF.js, and compatible with Google Cloud TPU infrastructure

---

## 🔬 Applications

- Advanced robotics and autonomous systems  
- Industrial visual inspection and surveillance  
- Healthcare AI in smart facilities  
- Cloud-scale computer vision pipelines  

---

## ⚙️ How It Works

```python
import tensorflow as tf

# Load the model
model = tf.saved_model.load("path/to/protoge-lg")

# Enable GPU/TPU if available
# Configure strategy if deploying on TPU

# Run full detection or provide custom labels
target_labels = ["robot arm", "conveyor belt", "monitor"]
detections = model(input_tensor, labels=target_labels)

# 🧪 Supported Modes
## 🔍 Full Mode
Detect and track all 1000+ categories in a single pass. Best used for exploratory environments or full-scene awareness.

## 🎯 Selective Mode
Pass in a list of labels to optimize speed and accuracy on known targets. Ideal for constrained detection environments like industrial automation or healthcare.

#📊 Performance Overview
Classes Supported =	 1000+
Acceleration	  =  GPU / TPU
Model Size	      =  ~60MB
Inference Speed	  <  40ms (TPU)
Video Tracking	  <= 60 FPS

## 📖 Citation

If you use **Celestial-Mini** in your work, please consider citing:

```
@misc{celestialmini2025,
  title={Celestial-Mini: A Lightweight Real-Time Object Detector},
  author={Lang, John},
  year={2025},
  howpublished={\url{https://huggingface.co/langutang/celestial-mini}}
}
```

---

## 📬 Contact & License

- 📫 For questions or collaboration, open an issue or contact the maintainer.
- ⚖️ License: MIT (see LICENSE file for details)

---

## 🌠 Hugging Face Model Hub

To load from Hugging Face:

```python
from transformers import AutoFeatureExtractor, TFModelForObjectDetection

model = TFModelForObjectDetection.from_pretrained("langutang/protege-lg")
extractor = AutoFeatureExtractor.from_pretrained("langutang/protege-lg")
```

---

Transform your edge AI projects with the power of **Protege-Lg** 🌠