## 🚀 **About This Challenge** 
####**What You Need to Do**

Participants will train a model on **synthetic images** generated by each individual on FalconCloud

**Note: Something different from last competitions is that we have provided a small number real-world images for the validation. These are for validation ONLY, not training.**

Then, the trained model will then accurately predict bounding boxes for **real-world images**.  
The model’s performance is evaluated based on [**mAP@50**.](https://falcon.duality.ai/secure/documentation/ex3-colab-testing-output?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1#the-meaning-of-50-in-m-ap-50)   

---

## **🛠 How to Get Started** 
[👉 Get started with FalconCloud](https://falcon.duality.ai/secure/scenarios/edit/7c5b51d0-553c-4bd3-9566-73723bbc43a5?utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
 
### Generate Initial Dataset(s)
1. [Go to the competition's page on the FalconCloud website for links to various resources, including the data-generating scenarios](https://falcon.duality.ai/secure/documentation/kaggle-page?utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1).
2. **[Start generating data](https://falcon.duality.ai/secure/scenarios/edit/7c5b51d0-553c-4bd3-9566-73723bbc43a5?utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)** using your personally selected variations. 
3. **Train a model of your choice (YOLO recommended)**
4. **Run predictions on test images** 
5. **Convert the labels in .txt to a submission.csv** with our `convert_to_csv.py` file and then submit the csv file.
 `python convert_preds_to_csv.py`
6. **Confirm the format of your predictions** are correct using the submission format described in the Evaluation Section.

### Analyze Model Weakness and Improve Datasets
1. Analyze and address weakness by generating more robust and targeted data using:
[The soup+ Multi-class data generating scenario.](https://falcon.duality.ai/secure/scenarios/edit/7c5b51d0-553c-4bd3-9566-73723bbc43a5?utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)

---

## **Workflow** 

1. Create a free Falcon account and run the FalconCloud Scenario
2. Configure the parameters you think will generate the most aligned data
    - [You can reference this page](https://falcon.duality.ai/secure/documentation/model-and-data-improvements?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1) for tips on effective synthetic data creation.
3. Generate multiple datasets using a variety of configurations, and identify dataset weaknesses
   - [Add occlusions](https://falcon.duality.ai/secure/documentation/colab-ex-1-whats-next?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
   - [Include similar twins for improving precision](https://falcon.duality.ai/secure/documentation/ex3-colab-whats-next?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
   - [Combine target datasets](https://falcon.duality.ai/secure/documentation/ex3-colab-whats-next?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)   
4. **Train a model** 
   - Local training instructions start on [this page](https://falcon.duality.ai/secure/documentation/ex3-training-installs?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
   - Google Colab cloud training instructions are on [this page](https://falcon.duality.ai/secure/documentation/ex3-colab-training-process?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
5. **Run predictions on test images** 
   - Local installation prediction instructions are on [this page](https://falcon.duality.ai/secure/documentation/ex3-test?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
   - Google Colab cloud prediction instructions are on [this page.](https://falcon.duality.ai/secure/documentation/ex3-colab-test?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
6. **Convert the labels in .txt to a submission.csv** with our `convert_to_csv.py` file and then submit the csv file.
 ```sh
python convert_preds_to_csv.py
```
7. **Confirm the format of your predictions** are correct using the submission format described in the Evaluation Section.
8. Create more data to target model weaknesses, align the synthetic data domain with the real-world domain, and create more effective training data.
9. Check the leaderboard to see how your model performs in comparison to other competitors!
10. **Join the [Discord Community](https://discord.com/invite/dualityfalconcommunity)** for solving queries related to the competition in the `kaggle-competitions` channels

### **Additional Option: Generate More Diverse Data Locally Using FalconEditor**   
1. **Sign up and download FalconEditor for free**: 👉 [Falcon EDU Account](https://falcon.duality.ai/secure/documentation/ex3-install?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)  
2. Complete **Exercise 3** to access a much larger range of variety and control.
   - [Click here to view computer restrictions under the "Support and Requirements" section.](https://falcon.duality.ai/secure/documentation/intro-to-the-exercises?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
   - [Click here to get started with FalconEditor on your local computer!](https://falcon.duality.ai/secure/documentation/ex3-objdetection-multiclass?sidebarMode=learn&utm_source=kaggle&utm_medium=kaggle_competition&utm_campaign=competition_fk1)
3. Improve model performance.

## 🎯 **Rules**
📌 **General Rules**
- **Teams**: Maximum **5 members** per team.
- **Pretrained Models**: You may use pretrained models.
- **Submission Limit**: 3 submissions per day.
- **Data**: You may ONLY use synthetic data from Falcon to train your model, but we have included some real-world images for the validation.

🚨 **Disqualifications**

Violations include but are not limited to:
- Manually modifying test labels.
- Using real world data for training.

🔗 **Leaderboard**
- **Public Leaderboard**: Uses a **50%** of the test set for real-time rankings.
- **Private Leaderboard**: Final ranking based on other **50%** hidden test labels.


---

## 📣 **Community & Support**
- **Join Duality's**  [Discord](https://discord.com/invite/dualityfalconcommunity) for Falcon help and troubleshooting. 
- **Follow us on LinkedIn**  for updates and opportunities📢
   - [Duality AI](https://www.linkedin.com/company/dualityai/)



💡 **Think you can build the best synthetic-to-real object detector? Let’s see what you've got!**