Open yingruiji opened 1 year ago
Thank you for your kind feedback! However, we just utilized these real-world datasets for cross-domain evaluation (in the zero-shot setting). If you intend to train models on these small datasets, you might need to reduce the model size and carefully adjust the learning rates to prevent overfitting.
SceneFlow dataset is on the servers and too big to send... Maybe you can try to download from official website(https://lmb.informatik.uni-freiburg.de/resources/datasets/SceneFlowDatasets.en.html) via BitTorrent with download tools like qbittorrent or 迅雷. Why downloading fails, is there any error information?
The data set always shows a progress of 0, and then after a while it shows that the download failed. Oh, yes, I noticed that in the mid14 dataset, if the directory read by pretrain is not included in the category directory, is it? The correct path is: ELFNet/MIDDLEBURY/testQ/categories/im0.png, actually: ELFNet/MIDDLEBURY/testQ/im0.png, I tried to modify it for a while but it didn’t seem to work, I wonder if you can update the dataloader part. Looking forward to your reply.
Maybe you need a VPN? I tried to download the dataset and no error occurs. As for Middlebury dataset, make sure your --dataset_directory
is /path_to_middlebury/trainingQ/
, which includes the subfolder name.
This is what I placed according to your data set : MIDDLEBURY | trainingQ | Motorcycle | disp0GT.pfm | disp1GT.pfm | im0.png | im1.png | mask0nocc.png | mask1nocc.png
But when I use pretrain to test middlebury.py, the error "category" cannot be recognized, and the path reminder is ELFNet/MIDDLEBURY/testQ/im0.png, But there is no error reported in kitti dataset.The subfile category "Motorcycle" was not recognized. Can you update the dataloader part in middlebury?
Hi ELFNet Team,
You did a great job making this project a reality! I noticed that you mentioned the training parameters of Scene Flow FlyingThings3D subset in your paper. I would like to know more about KITTI 2012 and KITTI 2015, and the training parameters of Middlebury 2014 dataset. If possible, can you share the script training parameters related to these datasets? Looking forward to your answer.
best wishes, Yingrui