Question: Are thresholds in llm_answer_hm3d/mp3d.txt LLM-generated or manually corrected?

[cOo2-Ma]

I noticed that the thresholds for the same items in the two files, /media/mkk/ApexNav/llm/answers/llm_answer_hm3d.txt and /media/mkk/ApexNav/llm/answers/llm_answer_mp3d.txt, are different. Could you please confirm if these were all generated by the LLM, or if they underwent manual correction?

[Zager-Zhang]

Hi! Yes, I can confirm that these results were all directly generated by the LLM at the time, without any manual correction.

Regarding the different thresholds between the two files, as explained in the appendix of our paper: For the MP3D dataset, some labels are uncommon, and certain scene and object shapes are incomplete, making it difficult for the detector to identify the targets. Therefore, we specifically adjusted the confidence threshold range to [0.25, 0.35] for MP3D to account for these harder-to-detect objects and adapt to the dataset's characteristics.

As for the differences in the specific items generated by the LLM, there are two main reasons:

1. Randomness parameters: The generation process of LLMs is inherently non-deterministic (influenced by parameters like temperature), meaning the specific outputs can vary slightly with each run.
2. Version iteration: LLMs are updated and iterated upon very quickly. Outputs will naturally differ depending on the specific version or time period the model was used.

Hope this helps clarify your question!