Details on Fine-Tuning for the Sim to Real Experiment

[AmandineBtto]

Hello,

Thank you so much for releasing your code! It's incredibly helpful.

I have a few questions regarding the fine-tuning procedure used for the "hearing anything anywhere" dataset, as I would like to replicate and better understand your approach:

Fine-tuning Procedure:

• Could you provide more details on the fine-tuning process?
• Are any layers frozen during fine-tuning? If so, which ones?
• What learning rate did you used? For how many steps?

Dataset Handling:
The "hearing anything anywhere" dataset does not contain panoramic depth maps. How do you handle the geometric context in the absence of this data? Is there a particular method or workaround you’ve implemented?

Room Configurations:
Are you using only the "base" room configurations, or do you also incorporate the other room variations such as Translation, Rotation, or Panel?

Reproducing Results:
Is there any additional information, configuration setting, or insight that would be helpful for successfully reproducing your results?

I really appreciate any clarifications you can provide!

[DragonLiu1995]

Sorry for the late reply, and thank you for your interest in the work! Regarding your question:

1. We use full weights finetuning, and use 8-shot reference RIRs sampled from12 observations to predict one of the other 4 RIRs each iteration, and then use the validation set as reference. The held out ratio is exactly same as Diff-RIR work.

2. You can get approximated mesh (planar geometry provided by HAA github repo) from the rooms in hearing anything anywhere dataset, and then render panorama depth map inside the room at any location.

3. We use the only base room configuration, other variations are not considered due to time limitations.

4. We compare the model against other methods at different sampling rate and RIR length (see paper for details)