Ques about table 5 and table 6?

[huixiancheng]

Hello, thank you for open-sourcing the relevant code. It's a very interesting and well-engineered experiment.

I'm not sure if you've noticed or analyzed the differences between Table 5 and Table 6. Since the two datasets are very similar, the results in terms of mIoU show a consistent trend. However, for PVT v3, the trend doesn't seem to be quite as consistent. Have you looked further into this?

Also, I noticed that both your code and paper mention LitePT, but there doesn't seem to be corresponding experimental results introduced in the paper. Would you be willing to share your relevant experience or results? I'm currently testing LitePT on datasets like Toronto-3D and would be very interested to know more.

[maxim0815]

Hey, thanks for your interest. Glad to hear.
Are you referring to the linear increase across all models and datasets, excluding PTv3? For the Toronto3D dataset, intensity was included as a feature channel to enable the detection of road markings. This was also the class where the model struggled the most.

We did not run experiments with it, as our experiments were already completed when LitePT was published 15.12. However, we added it to the repository because it is faster and supports our main goal of achieving a large receptive field with reduced memory usage. Unfortunately, we have not yet trained it with different setups, sorry.