We would appreciate some clarifications regarding rule 3.6.3 (“Violations”), specifically in relation to computer vision:
- Any software specifically designed to complete any single primary task of RoboCupJunior Rescue,
e.g., any letter recognition libraries such as Tesseract or EasyOCR, etc., will be prohibited.
We are relatively new to computer vision techniques and may not have a complete understanding of the subject. We want to ensure clear communication and appreciate your help in finding a common language.
We are diligently working to interpret this rule objectively, but we seek further clarification on its boundaries. We are concerned about unintentional rule violations while also aiming to avoid reinventing the wheel.
- Can we utilize an existing model, such as those available in thetf2 model zoo, and train it from scratch using our own dataset?
- If using an existing architecture is not allowed (as creating a new one is inconceivable for the competition’s objectives), would we be required to implement functions like model fitting ourselves?
- Are we responsible for creating loss functions as well, if we need to develop our own architecture?
- If the use of a model from the Model Zoo is permitted, are other real-time object detection systems also allowed?
To clarify further, we do not intend to use an existing pre-trained object detection model. We have prepared our own dataset, and the training phase will commence with random weights. However, we are struggling to determine the specific boundary defined by the rule.
We appreciate any assistance you can provide in helping us.