❓ Frequently Asked Questions (FAQ)

[mattheliu]

❓ Frequently Asked Questions (FAQ)

This post will be updated as new questions come in. If your question isn't answered here, feel free to reply below!

---

Topic Selection

Q: Can I choose any OCR-related topic, or is it limited to the examples in the rules?

A: Completely open. The scenario pool in the rules is for reference only — you are encouraged to define your own task direction. The key criteria are: real-world need, scenario scarcity, and that existing OCR methods don't handle it well.

Q: What counts as a "long-tail" scenario?

A: Scenarios that are underserved by current academic benchmarks and commercial solutions. Examples include minority language recognition (Tibetan, Arabic), handwritten forms, organic chemical formula recognition, flowchart parsing, and medical prescription recognition. Common scenarios like standard text or table recognition are not recommended as they already have mature solutions.

Q: Can I work on a scenario that another participant has chosen?

A: Yes. There are no exclusive topic claims. Different approaches to the same scenario can both be evaluated on their own merits.

---

Dataset & Evaluation Set

Q: What is the evaluation set, and why does it have an earlier deadline (Apr 24)?

A: The evaluation set is a held-out benchmark you construct for your chosen scenario — it's used by reviewers to objectively assess model performance. It has an earlier deadline because the review cycle requires time to process all submissions before the preliminary results (May 25).

Q: What should the evaluation set include?

A: At minimum: images/documents, annotations, a task description, and an evaluation script. You should also provide a dataset description covering data sources, scale, category distribution, and difficulty analysis. Host it on Baidu Netdisk or AI Studio Open Datasets and submit the link via email.

Q: Does the training data need to be open-sourced?

A: No. Only the evaluation set link (submitted via email) and the open-source project (GitHub + Hugging Face) are required to be shared. Training data is not required to be public.

---

Model & Fine-Tuning

Q: Which base models can I use?

A: PaddleOCR-VL series models (e.g., PaddleOCR-VL, PaddleOCR-VL-1.5). Refer to the fine-tuning guides:

• PaddleOCR-VL Best Practices
• PaddleOCR-VL-1.5 Best Practices

Q: Is computing resources provided?

A: No. Participants need to provide their own compute. All participants who submit a project link will receive RMB 100 GPU credits from PaddlePaddle AI Studio to help with tuning. Lightweight models or quantization techniques are recommended for the preliminary round.

Q: Can I use synthetic data for training?

A: Yes. Data collection methods including synthesis are explicitly listed as part of the training data construction report requirements.

---

Submission & Rules

Q: How do I submit my work?

A: All materials (except finals defense slides) are submitted via email to ext_paddle_oss@baidu.com, with subject: PaddleOCR Derivative Model Challenge - [Material Name] - [GitHub ID].

Q: Can I update my submission before the deadline?

A: Yes. All materials can be iterated continuously before their respective deadlines. The final version is determined by the last submission before the deadline.

Q: Can Baidu employees participate?

A: Yes, Baidu employees can participate and receive rankings and honors, but per company policy they do not receive cash prizes.

Q: Can a team have more than one person?

A: Yes, up to 5 members per team.

---

Scoring & Prizes

Q: Are the preliminary and finals scores cumulative?

A: No. Finals rankings are re-evaluated entirely based on finals submissions. Preliminary scores are not carried over.

Q: Can I win both the Data Contribution Award and a finals prize?

A: Yes. The High-Quality Evaluation Set Contribution Award (RMB 1,000 × 10 people) and finals prizes can be awarded cumulatively.

---

Last updated: April 2025 | Maintained by @liu-jiaxuan @cuicheng01