search

ywyue / RoomFormer

[CVPR 2023] RoomFormer: Two-level Queries for Single-stage Floorplan Reconstruction
https://ywyue.github.io/RoomFormer/
MIT License
163 stars 22 forks source link

Train on my dataset #19

Closed yyyyll0ss closed 9 months ago

yyyyll0ss commented 9 months ago

Thank you for your open source work. When I tried to use my own three-channel image data set, it seemed that I could not train, there were no bugs in the training, and the loss was constantly decreasing, but the accuracy of the valuation set was always 0. Do you know the possible reasons for this?

ywyue commented 9 months ago

Thank you for your interest in our work. Based on the currently provided information, I can not judge the reasons - could you first try to plot the predictions and also ground truth to see whether they are reasonable? If the loss was decreasing but the accuracy was always 0, then it is very likely that there are some issues with the evaluation.

ywyue commented 9 months ago

Closed this issue for now. Feel free to let me know if you have further concerns.

claumeh commented 3 weeks ago

Can you share a sample density map from an independent dataset, so we can get a better idea on the prediction for different data? I would like to run the eval on this density map. Thanks

ywyue commented 3 weeks ago

hi @claumeh , what do you mean by "independent dataset"? If you already have your own 3D scan, it is easy to project the 3D point cloud to a density map. The projection code can also be found in this repo.

claumeh commented 2 weeks ago

hi @ywyue , here is a density map generated from my own dataset. But the results are not very accurate in all evaluation scenarios:

Can you try to predict better results from this density map? 03268