Johan Edstedt

https://github.com/ScanNet/ScanNet/issues/28 I guess the issue is that I first need to adjust by the extrinsics? Perhaps this could be made more clear?

Although it seems like they have already been aligned in this way? I don't understand how they can be so wrong after postprocessing?

Bump

We do not plan to release the training code in the near future, however, if you have any questions regarding the training I'm happy to answer. Probably the training code...

@cvbird 1. In the current version (a new preprint will be released soon) we train either on only megadepth (like LoFTR), or on a combination of MegaDepth and a synthetic...

@KakueiTanaka Hi, we use height = 384 width=512 images, we will release preliminary training code, loss functions etc in a few days :)

@KakueiTanaka I updated the codebase now, and there is some (hopefully working) training code. Note that the code we provide here is adapted from our internal training framework that is...

I realized that in our internal training we actually freeze the batchnorm statistics in the resnet backbone during training, which is not done here. This might cause some small discrepancies...

I think this problem is also prevalent in multiple other parts of kornia, where the coordinate systems used are not explicit. create_meshgrid e.g. assumes a [0,0] center of top-left pixel,...

@edgarriba I'm pretty sure that `align_corners = False` implies [0.5,0.5] center of top-left. I've had so many issues with this when doing geometric matching haha.