# Competition Rules — VinFast Car Detection (mAP@0.5)

By participating in this competition, you agree to comply with all rules below. The organizers may disqualify submissions that violate these rules.

---

## 1. Eligibility
- This is an open competition for learning and experimentation.
- You must use a single Kaggle account. Multiple accounts for the same participant are not allowed.
- Teams are allowed unless explicitly disabled by the organizer.

---

## 2. Competition Objective
- Build an **object detection model** to detect **VinFast cars (1 class)** in images.
- Your submission will be scored using **mAP@0.5**.

---

## 3. Submission Limits & Leaderboard
- **Daily submission limit:** up to **5 submissions per day** per participant/team.
- Submissions are scored on the **public leaderboard** during the competition.
- Final ranking is determined by a **private test split** and shown on the **private leaderboard** after the competition ends.

---

## 4. Allowed Data & External Resources
### ✅ Allowed
- The provided competition dataset (train images + labels).
- **Pretrained weights** (e.g., COCO/ImageNet).
- Publicly available libraries, code frameworks, and tutorials (e.g., PyTorch, OpenCV).
- Any feature engineering, augmentation, pseudo-labeling, and ensembling **as long as it only uses the provided images**.

### ❌ Not Allowed (to keep the competition fair)
- Using **external labeled datasets** for VinFast detection or vehicle bounding boxes that are not part of this competition data.
- Using any additional VinFast-labeled data collected from the internet or private sources.
- Manually labeling/annotating the hidden test images.
- Any method that attempts to infer the hidden test labels directly (label leakage / reverse engineering).

**Clarification:** Pretrained models are allowed, but you must not train on additional external **annotated** datasets that overlap the target task.

---

## 5. Inference Requirements
- You may run inference and training on your own machine or in Kaggle Notebooks.
- You may use ensembling, TTA (test-time augmentation), and post-processing (NMS/Soft-NMS/WBF).
- Do not hardcode predictions for specific test images.

---

## 6. Reproducibility & Sharing
- Sharing code and notebooks is encouraged.
- If you publish a solution, please include:
  - training configuration (model, imgsz, epochs, augmentation)
  - post-processing (confidence threshold, NMS settings)
  - how you generated the final submission file

---

## 7. Integrity & Fair Play
- Do not exploit bugs, unintended data access, or loopholes in the evaluation system.
- Do not scrape or attempt to identify the test labels through metadata, filenames, or hidden directories.
- Any cheating, collusion, or unauthorized access will result in disqualification.

---

## 8. Content & Privacy
- Do not upload personally identifiable information (PII) in discussions.
- Do not upload the hidden test labels or any derived ground truth.

---

## 9. Prize / Ranking (if applicable)
- Rankings are determined solely by the private leaderboard score (mAP@0.5).
- In case of ties, the organizer may use:
  1) earliest submission time achieving the tied score, or  
  2) additional hidden evaluation images (if available).

---

## 10. Organizers’ Rights
- The organizer reserves the right to:
  - update rules for clarity and fairness
  - remove invalid submissions
  - disqualify participants who violate rules
  - end the competition early in case of abuse

---

## 11. Attribution & License
- Dataset is provided under **CC BY 4.0**.  
  Please give attribution when publishing derived works, notebooks, or models.

Original source (Roboflow Universe):
https://universe.roboflow.com/dsp-project/vin-detection-hqtw7

---

## 12. Questions
Use the **Discussion** tab for questions about:
- submission format
- metric interpretation
- baseline training tips
- common errors