Create README.md

# 🍎 ApplesM5-Dataset

This dataset contains annotated object detection data used in the **Synetic AI Apple Benchmark** study, measuring the effectiveness of rendered (synthetic) data versus real-world data for training small vision models. The dataset was constructed using photorealistic, physics-accurate 3D renders of apples in orchard scenes, with perfect annotations and environmental diversity.

This dataset supports object detection models such as YOLOv8 and RT-DETR, and includes:

- RGB images
- Bounding box annotations (COCO format)
- Real and rendered training/validation splits
- Metadata for benchmark reproduction

## 📊 Use Cases

- Object detection
- Real vs synthetic data performance evaluation
- Model training and validation
- Benchmarking data efficiency in agriculture

## 📁 Dataset Structure
ApplesM5-Dataset/
├── images/
│ ├── train/
│ ├── val/
├── annotations/
│ ├── instances_train.json
│ ├── instances_val.json
├── metadata/
│ ├── image_metadata.csv


## 🔬 Benchmark Context

This dataset was used in the Synetic AI whitepaper to compare multiple training strategies:
- Real-only
- Rendered-only
- Rendered + real validation
- Joint (rendered + real) training

Rendered data outperformed real data by up to **34% mAP** in certain configurations, especially at low confidence thresholds where operational reliability matters most.

## 📄 Citation & Whitepaper

🔗 Whitepaper coming soon. Visit [synetic.ai](https://synetic.ai) for updates.

Once published, this section will include the official citation and DOI link.

## 🔧 License

MIT License

## 🔤 Language

No language data. This dataset is image-based.

## 🏷️ Tags

`synthetic-data`, `object-detection`, `agriculture`, `benchmark`, `rendered`, `real-vs-synthetic`

## 🎯 Task Categories

- Object Detection

## 📦 Size Category

10K < # images < 100K

---

**Contact**: For questions or commercial licensing, please visit [synetic.ai](https://synetic.ai).

README.md

@@ -1,110 +1,19 @@
-# ApplesM5
-## Breaking the Bottleneck: Synthetic Data as the New Foundation for Vision AI
-
-This repository contains training images and scripts for the Synetic AI **ApplesM5** object detection project that was used in the **Breaking the Bottleneck: Synthetic Data as the New Foundation for Vision AI** white paper, using **Ultralytics YOLO12**. The core scripts allow you to train models using custom YAML datasets and evaluate the results using the provided train_metrics.py script.
-
-The paper is available for download at https://synetic.ai/white-paper/breaking/benchmark .
-
----
-
-## 📂 Repository Structure (Key Files)
-
-| File                       | Purpose                                                                                     |
-| -------------------------- | ------------------------------------------------------------------------------------------- |
-| `PrepareDatasets.py`       | Produces the individual datasets used for training various combinations.                    |
-| `applesm5-train-det.py`    | Trains YOLO12 detection models using specified datasets and hyperparameters.                |
-| `FileCrawler.py`           | Recursively crawls directories to find image and label files. Used for evaluating datasets. |
-| `train_metrics.py`         | Runs evaluations on trained YOLO12 models and computes mAP, precision, and recall metrics.  |
-| `*.yaml` (dataset configs) | Define dataset splits, including training, validation, and test image directories.          |
-
----
-
-## ⚙️ Setup
-
-### 1. Install Dependencies
-
-```bash
-pip install ultralytics tqdm
-```
-
-Your environment should have **PyTorch** and GPU drivers properly configured.
-
----
-
-## 🚀 Usage
-
-### 0. Prepare Datasets (`PrepareDatasets.py`)
-
-It will produce multiple folders combinations of synetic and real from the real and synetic source folders.
-
-```bash
-python PrepareDatasets.py
-```
-
-### A. Training Models (`applesm5-train-det.py`)
-
-To train object detection models using YOLO12:
-
-```bash
-python applesm5-train-det.py
-```
-
-Key things to configure:
-
-- Edit the `dataNames` list to point to your dataset YAML files (e.g., `real`, `synetic+real`, etc.).
-- YAML files should be placed at `/home/user/datasets/ApplesM5/`.
-- Adjust `hyperparams`, `epochs`, and GPU `devices` as needed.
-- The script trains multiple model variants (`yolo12n.yaml`, etc.) and saves results to the Ultralytics default `runs/detect/` folder.
-
----
-
-### B. Dataset YAML Files
-
-Example dataset YAML (`real.yaml`):
-
-```yaml
-path: /path/to/your/dataset
-train: images/train
-val: images/val
-test: images/test
-names:
-  0: apple
-```
-
-Modify the paths in your YAML files to point to your dataset locations.
-
----
-
-### C. Evaluating Models (`train_metrics.py`)
-
-After training, you can evaluate your models on a validation dataset:
-
-```bash
-python train_metrics.py
-```
-
-Make sure to adjust the following in the script:
-
-- `modelPaths`: list of trained YOLO12 model `.pt` files to evaluate.
-- `pathValsDataset`: path to your validation images (`.png`/`.jpg`).
-
-This will compute **mAP50**, **mAP50-95**, **precision**, and **recall** scores and print them to the console.
-
----
-
-## ✅ Example Workflow
-
-1. Prepare datasets and YAML config files.
-2. Train detection models with `applesm5-train-det.py`.
-3. Run `train_metrics.py` to benchmark models.
-4. Iterate on your datasets and training parameters to improve performance.
-
----
-
-## 🔧 Notes
-
-- The training script assumes a multi-GPU setup (adjust the `devices` list if needed).
-- The repo is tuned for an NVIDIA DGX or similar system with 8 GPUs but can be modified for single-GPU setups.
-- Dataset YAML and trained model `.pt` files follow the **Ultralytics YOLO12** conventions.
-
 ---
+license: cc-by-4.0
+task_categories:
+- object-detection
+language:
+- en
+tags:
+- synthetic-data
+- apple-detection
+- computer-vision
+- photorealistic
+- benchmark
+- agriculture
+- yolo
+- object-detection
+pretty_name: 'Better Than Real: Synthetic Apple Detection for Orchards'
+size_categories:
+- 1K<n<10K
+---