Transductive pseudo-labeling / self-training on the test images is allowed, provided the entire process is fully automated.

To ensure fairness and transparency across all participants, please adhere to the following mandatory rules regarding pseudo-labeling for this and future challenges:

• Fully Automated Process: The generation of pseudo-labels must be strictly algorithmic. You may use your trained models to infer labels on the test set and fold all or part of those predictions back into your training process. No manual annotations, human-in-the-loop visual transcription, or manual labeling of test images is permitted under any circumstances.
• Rules on Post-Processing: Automated post-processing rules applied to generated pseudo-labels are permitted - for example, programmatic auto-corrections based on domain logic (e.g., automatically replacing a predicted numeric 0 with the letter O in specific word contexts). However, these rules must be generalisable and code-driven, not hardcoded fixes designed to target specific individual test images observed manually.
• Code Submission & Reproducibility: If your final solution utilizes pseudo-labeling, the entire pseudo-label generation pipeline (including inference, post-processing logic, and re-training loops) must be clearly explained in your solution documentation and fully reproducible via your submitted code.