---
license: mit
library_name: ultralytics
pipeline_tag: object-detection
base_model: Ultralytics/YOLO11
base_model_relation: finetune
language:
  - en
tags:
  - ultralytics
  - yolo
  - yolo11
  - yolo11n
  - yolov11
  - yolov11n
  - object-detection
  - computer-vision
  - waste-detection
  - trash-detection
  - garbage-detection
  - recycling
  - recycling-automation
  - waste-sorting
  - edge-ai
  - onnx
  - tflite
datasets:
  - dmedhi/garbage-image-classification-detection
  - garythung/trashnet
---

# MIRA - YOLO11n Waste Detection for Recycling and Waste Sorting

MIRA is a custom YOLO11n object-detection project for waste detection,
recycling automation, and automated waste sorting.

The models detect five classes:

- glass
- metal
- paper
- plastic
- trash

This repository contains PyTorch, ONNX, and TFLite exports from the MIRA
experiments. The recommended reference model is `mira_exp019.pt`.

- GitHub: https://github.com/jeremy341/MIRA-AI
- Project website: https://mira-vision.vercel.app/
- PyPI package: https://pypi.org/project/mira-ai/

## EXP-019 performance

| Metric | Result |
|---|---:|
| mAP50 | 90.58% |
| mAP50-95 | 82.15% |
| Precision | 87.2% |
| Recall | 84.6% |
| Training images | 5,108 |
| Validation images | 415 |
| Test images | 1,375 |
| Classes | 5 |

These results come from the documented evaluation split used for EXP-019.

## Available model files

| File | Format | Description |
|---|---|---|
| `mira_exp019.pt` | PyTorch | Recommended YOLO11n detector |
| `mira_exp019.onnx` | ONNX | ONNX export of EXP-019 |
| `mira_exp019_int8_320.tflite` | TFLite | INT8 export at 320 px |
| `mira_exp019_int8_640.tflite` | TFLite | INT8 export at 640 px |

The repository also contains models from earlier MIRA experiments.

## Experiment results

| Experiment | Model | Dataset | mAP50 |
|---|---|---|---:|
| EXP-005 | YOLOv8n | Custom + TrashNet | 82.3% |
| EXP-006 | YOLOv8n | Fused Wild + TrashNet | 39.4% |
| EXP-009 | YOLOv8n | TrashNet | 72.8% |
| EXP-011 | YOLOv8n | TACO | 35.0% |
| EXP-013 | YOLO11n | TACO + TrashNet | 55.1% |
| EXP-014 | YOLO11n | Combined dataset | 60.7% |
| EXP-015 | YOLO11n | Combined dataset with WaRP | 56.0% |
| EXP-016 | YOLO11n | WaRP-focused dataset | 58.8% |
| EXP-017 | YOLO11n | Larger combined dataset | 59.3% |
| EXP-018 | YOLO11n | Clean balanced dataset | 90.6% |
| EXP-019 | YOLO11n | Clean balanced repeatability run | 90.58% |

The main lesson was that adding more data did not automatically improve the
model. Removing inconsistent examples and building a cleaner, more balanced
dataset led to the strongest results in EXP-018 and EXP-019.

## Quick start

Install the required packages:

```bash
pip install ultralytics huggingface_hub
```

Download the recommended model directly from Hugging Face:

```python
from huggingface_hub import hf_hub_download
from ultralytics import YOLO

model_path = hf_hub_download(
    repo_id="Jeremy341/MIRA-AI",
    filename="mira_exp019.pt",
)

model = YOLO(model_path)

results = model.predict(
    "image.jpg",
    conf=0.25,
    save=True,
)

results[0].show()
```

For validation, provide a compatible YOLO dataset configuration:

```python
results = model.val(data="dataset.yaml")
```

## Intended use

MIRA is intended for research and prototyping in:

- waste detection
- recycling automation
- waste sorting
- computer-vision research
- edge-AI object detection
- robotic sorting experiments

The models are not presented as a finished production recycling system.

## Datasets

The models were trained using combinations of:

- [dmedhi garbage image classification/detection](https://huggingface.co/datasets/dmedhi/garbage-image-classification-detection)
- [TACO](https://github.com/pedropro/TACO)
- [TrashNet](https://github.com/garythung/trashnet)
- [Roboflow Trash Detection](https://universe.roboflow.com/jerry-jukbu/trash-detection-1fjjc-uqlv1)

The datasets were remapped to the five MIRA classes. Each dataset remains
subject to its original license and usage terms.

## Limitations

The models can struggle with:

- white crumpled paper
- cans viewed from the opening
- strongly overlapping objects
- unusual lighting
- unusual viewing angles
- waste objects outside the training distribution

The reported results do not guarantee the same performance on completely
independent real-world images.

## License

The model files are provided under the MIT License where applicable. Dataset
licenses remain subject to their original terms.