XingliangJin
commented
3 months ago First, thank you for following our work!
-
Your finding indeed warrants further exploration. I have rechecked the accuracy of my previous experiments, and they align with your results. In fact, when training the model, I did not specifically focus on generator accuracy because the novelty of my paper primarily lies in the fusion of style and content semantics. The stylization framework directly utilizes the work from DeepMotionEditing. You could retrain this project to see if their accuracy meets your expectations. From my perspective, the reason for the observed low accuracy might be that the discriminator was trained on a single style label, while our generator, by integrating semantics, refines the style to be related to semantics rather than maintaining a uniform style. During training of FineStyle, the generator may not be significantly influenced by the discriminator, leading to low accuracy because high accuracy would mean our generator would only produce a uniform, unrefined style. In other words, the fine-grained style generated by our method results in lower accuracy when judged by a discriminator trained on a single style. This suggests two things: firstly, our FineStyle might not require a discriminator; secondly, enhancing the discriminator with motion semantics might improve its effectiveness, potentially increasing the generator's accuracy.
-
The currently open-sourced version does not include the evaluation code due to time constraints at that time, and it has been a long time since then, so it might not be feasible to provide the code directly now. However, the evaluation code primarily relies on this project: Evaluation-for-Motion-Puzzle. Additionally, some of the latest style transfer works have open-sourced their evaluation codes on the Xia dataset. The principle remains the same, mainly involving training classifiers and calculating the style or content accuracy of generated motions.
If you have any other questions or need more detailed information, please let me know.
I hope this message finds you well. I am currently working with your FineStyle code as part of my research. I have encountered a couple of questions regarding the implementation and performance of the model, and I would greatly appreciate your guidance.
Generator Accuracy Issue: After training the model for 300,000 iterations without changing the configuration file, I observed that the generator's loss has generally decreased, which is expected. However, the generator's accuracy remains very low, around 0.0X. Is this expected behavior, or could it indicate an issue with the training process or the configuration? I am particularly concerned that such low accuracy might reflect a fundamental problem with how the generator is learning or interacting with the discriminator.
Quantitative Evaluation Metrics: The paper mentions three key evaluation metrics: Frechet Motion Distance (FMD), Content Recognition Accuracy (CRA), and Style Recognition Accuracy (SRA). I am trying to locate the specific methods or classes in the code that implement the calculation of these metrics. Could you please guide me to the exact locations in the code where these metrics are computed? Understanding this will greatly help in validating the performance and results of the model.
Thank you very much for your time and assistance. Your work on FineStyle is truly impressive, and I am eager to ensure I am implementing and evaluating the model correctly.