The principle for selecting these polygon points

[shenhai911]

Thank you for your excellent work. I would like to ask, In these codes, what's the principle for selecting these polygon points? Is it necessary? Does this principle also apply to other datasets, such as KITTI and MulRan?

[shenhai911]

Hello, there are two more questions to ask:

1. the paper mentions that 20,167 frames of training submps and 17,582 frames of testing submaps can be obtained from sequences 00 and 01. However, using the provided codes, only 19,096 frames of training submaps and 7,420 frames of testing submaps are obtained from sequences 00 and 01. What are the differences between the relevant parameters in the paper and these codes? (1) results in the paper:

   (2) results from the provided codes

2. Regarding the "_OFFSET" in the code, why is it only used when processing the Karawatha dataset, but not when processing the Venman dataset?

[JBKnights]

Hi Shenhai911, to address your questions in order:

1. The training and testing regions were qualitatively selected to cover a diverse range of environment types within the dataset, such as wider open areas, more narrow and constrained paths, etc, for the purpose of having a more representative inter-sequence evaluation. In terms of how this relates to the KITTI or MulRan dataset, I would hazard that if one were to construct an inter-sequence testing setup then ideally the regions held out for testing queries should be representative of the diversity found within the sequences as well, though the nature of that diversity would obviously differ for an urban environment versus a non-urban environment.

2. For clarification, where in the paper did you read the numbers 20,167 and 17,582 from? The number of training maps in the paper (as you've captured in a screenshot here) seems to match the output from the code you've also shared, so I'm not sure where these numbers come from. I believe that the number of testing submaps quoted in the paper should also include submaps for the intra-sequence evaluation, which would include all the submaps from sequence 03 and 04 in both environments.

3. The _OFFSET value is used to separate the ground truth positions of Venman and Karawatha, since both have the starting point of their environments centered around (0,0,0); without the offset, it's possible during training for submaps from Venman and Karawatha to select each other as possible examples, which for obvious reasons is undesirable behaviour.