# Grocery Items: Multi-class Object Detection — Data

来源：https://www.kaggle.com/competitions/grocery-items-multi-class-object-detection/data

采集时间：2026-10-19T07:23:34Z

# Dataset Description

## 📁 Files Provided

```plaintext
/Starter_Dataset
├── /testImages/                        # Real-world test images (you'll be judged on)
│   └── /images/                       
├── /Dataset/                           # Synthetic data to start training, create more on FalconCloud
│   ├── /train/
│   │   ├── /images/  
│   │   └── /labels/
│   └── /val/                               #Synthetic and real-world data for validation, create more synthetic on FalconCloud
│       ├── /images/
│       └── /labels/

├── convert_preds_to_csv.py            # Script to convert YOLO predictions to Kaggle CSV
├── predict.py                         # YOLOv8 inference script
├── yolo_params.yaml                   # Configuration file for training/inference
```

### 🔑 Key Files:
- **Starter_Dataset.zip**: Includes synthetic training/validation data folders and the real-world test set.
- Dataset contains:
  - `/train/images` and `/train/labels` – for training
  - `/val/images` and `/val/labels` – for validation # includes real-world data ONLY for validation, not training
- `/TestImages/images`: Contains real-world test images used for scoring.
- **classes.txt**: Defines class names (`cheerios`, `soup`).
- **yolo_params.yaml**: Used by training and prediction scripts to set dataset paths and options.
- **convert_preds_to_csv.py**: Converts YOLO prediction `.txt` files into Kaggle submission format.
- **predict.py**: Creates the predictions for scoring

---

## 📊 Data Format

### 1. **Images**  
- Format: `.jpg` or `.png`  
- Resolution: All synthetic images are `1920x1080px` , some portrait and some landscape
- Located in `/images/` subfolder

### 2. **Labels (YOLO Format)**  
Each image has a corresponding `.txt` file inside the matching `/labels/` folder.  
Each line in the file follows this format:
- class_id confidence x_center y_center width height
- Coordinates are **normalized** (0 to 1) relative to image dimensions  
- Example (`IMG_0001.txt`): 
a. Single object: 0 0.95 0.45 0.50 0.20 0.15
b. Two objects: 1 0.95 0.45 0.50 0.20 0.15 0 0.92 0.70 0.30 0.10 0.12

### 3. **Class ID**  
-  **class 0**: cheerios,  **class 1**: soup , **class 2**: candles
- There are no other object categories in this challenge.

---

## 💡**IMPORTANT!** 
 
The **Scenario set up on Falcon Cloud** offers a way for you to generate a **richer set of high-variability synthetic images**, for helping **improve model generalization** and **real-world performance.**  

Leverage the tool and create your own targeted datasets for better results and a competitive edge on the leaderboard!

📥 [Start generating your own data here!](https://falcon.duality.ai/secure/scenarios/edit/7c5b51d0-553c-4bd3-9566-73723bbc43a5)

---

## 🎯 What You're Predicting
Predict **bounding boxes** for real-world test images in YOLO format:  
- csv file of all predictions with 2 major entries being, "image_id", "prediction_string"  
- Each line: `image_id", " object_class confidence x_center y_center width height"`  
- **Confidence scores** must be between 0 and 1

---

## 📖 Acronyms
| Term | Definition |  
|------|------------|  
| **YOLO** | You Only Look Once (object detection architecture) |  
| **mAP@50** | Mean Average Precision at 50% IoU |  
| **IoU** | Intersection over Union (box overlap metric) |  

---

## ❓ Frequently Asked Questions

**Q: Can I use external data for training?**  
A: Only if it's created using FalconEditor or FalconCloud.  

**Q: What image sizes and formats should I expect?**  
A: Synthetic images are 1980x1080px in .png format. Real test images will be 1920x1080px. 

**Q: How should I improve model performance and results beyond a certain point?** 
A: [Check out our tips and tricks page](https://falcon.duality.ai/secure/documentation/model-and-data-improvements?sidebarMode=learn?utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk_1) for improving performance through scenario changes, data manipulation, model tuning, and more!