Jianyuan Wang
Jianyuan Wang
Hi all, as stated in ReadMe, I would reorganize the codes in May to check if we can avoid such a problem for some users.
It would (possibly) involve an alternative RANSAC implementation to prevent the CUDA problem.
Hi, You may refer to [this](https://github.com/weixk2015/DeepSFM/blob/393ad8f39c5a305f3b351af497dc510697ee931e/sequence_folders.py) from DeepSfM for the Demon loader. Best
After checking the components, I think it may be ok to replace the third, fourth, and fifth frames by the image of the second frame, because DeepV2D takes an argmax...
Hi, have you solved this problem? I am also wondering how to obtain the ground truth optical flow
Hi @prajnan93 , Thank you for the interest! FlyingChairs dataset has 22232 training images, and we train the model with a batch size with 64 on it, as shown on...
Hi, Yes, the overall process is like this. Please notice that the learning rate should be reduced to half of the value used by the original PWCNet to keep the...
Hi, We could install minieigen by pip, like shown [here](https://package.wiki/minieigen). It should work with 0.5+, and I would double check the version later. By the way, minieigen is only used...
You can also refer to [this](https://github.com/weixk2015/DeepSFM/issues/8), which lists some possible solutions for this problem, but could be kind of complicated.
Just depth information is necessary. You can follow the codes of sunrgbd dataset loader https://github.com/facebookresearch/votenet/blob/master/sunrgbd/sunrgbd_detection_dataset.py.