Questions about Error Injection during Training #95
Description
Hello! Thank you for the excellent work and for open-sourcing the code.
After reading the paper and the codebase, I have a question regarding the training. I noticed that the selection of the error during training is completely random.
Because of this randomness, it seems highly likely that an error extracted from one domain (e.g., a landscape video) might be injected into a video from a completely different domain (e.g., a video about animals or humans).
Wouldn't this content inconsistency between the source of the error and the target video cause issues or introduce harmful noise during training? I am curious to know if the model is robust to this domain mismatch or if there is an underlying mechanism to handle it.
Looking forward to your insights. Thank you!