# Glove Labelling Model (SAM 2.1 Fine-Tuned)

This model is a fine-tuned [Segment Anything Model (SAM 2.1)](https://github.com/facebookresearch/segment-anything) designed specifically for **baseball glove segmentation**. It identifies fine-grained regions on a pitcher’s glove from video frames, with the goal of analyzing glove position, shape, and movement across pitches.

---

## 🔍 Model Details

- **Architecture**: SAM 2.1 Hiera-L variant
- **Framework**: PyTorch
- **Training Type**: Image-only fine-tuning on custom glove segmentation data
- **Losses**: Dice, IoU, and mask loss
- **Epochs**: 50
- **Batch Size**: 2
- **Dataset**: Custom COCO-format sequences of glove mask annotations split by pitch

---

## 🏷️ Labels (Classes)

This model supports six segmentation classes:
- `glove_outline`
- `webbing`
- `thumb`
- `palm_pocket`
- `hand`
- `glove_exterior`

---

## 📁 Files in This Repo

| File                  | Description                              |
|-----------------------|------------------------------------------|
| `pytorch_model.bin`   | Trained PyTorch weights (`.pt` file)     |
| `config.json`         | Model and dataset configuration          |
| `README.md`           | You're reading it                        |

---

## 🚀 Deployment Options

You can deploy this model using:
- **Google Cloud Vertex AI** (via Model Garden)
- **TorchServe**
- **CVAT** (via a custom segmentation model)
- **Hugging Face Inference Endpoints** (manual handler required)

---

## 🔗 Author

Created and maintained by [`caball21`](https://huggingface.co/caball21)  
Please cite if used in academic or production applications.