search

NVIDIA / semantic-segmentation

Nvidia Semantic Segmentation monorepo
BSD 3-Clause "New" or "Revised" License
1.77k stars 387 forks source link

How can I get the video data of CamVid? #22

Closed irfanICMLL closed 5 years ago

irfanICMLL commented 5 years ago

Thank you for your great work! Could you please help with the data pre-processing? How can I get the video data of Camvid? I only find some MXF video on the official database page, but the link to the code of pre-processing is broken.

Thanks, yifan

bryanyzhu commented 5 years ago

Hi, you can just use the data at here. They are already pre-processed. Many people follow its data structure, as do we. Hope this helps.

irfanICMLL commented 5 years ago

Thank you for your quick reply. The SegNet only has the scaled data with 480*360, and only the labeled ones. What is the source data you use to implement the aug? I thought there are some unlabeled video data like the sequence in Cityscapes.

bryanyzhu commented 5 years ago

Yes, I downloaded the videos from here as you mentioned, using ffmpeg to decode the videos into frames, doing some renaming and sanity check to get the unlabeled video data.

Since the official pre-processing code is broken, we don't know which frames are labeled. So I manually find the correspondence by watching the videos frame by frame. It takes some time but pretty straightforward.

irfanICMLL commented 5 years ago

Would you mind to share the unlabeled video data, so that we can reproduce your experiments?

Thanks!

bryanyzhu commented 5 years ago

I'm sorry but I don't have that data available. I left those when the internship ends. As I mentioned, decoding the videos here will give you the unlabeled video frames, you should have the same data as mine. Sorry about the inconvenience.

irfanICMLL commented 5 years ago

What a pity. By the way, have you used the pre-trained weight on cityscapes to train your CamVid model?

bryanyzhu commented 5 years ago

Yes, I used it. As shown in my experiments and previous VideoGCRF work, using pre-trained weight on Cityscapes to train CamVid is very helpful.

irfanICMLL commented 5 years ago

Thanks for your help.

giulionf commented 2 years ago

Seems like the project moved here: http://web4.cs.ucl.ac.uk/staff/g.brostow/MotionSegRecData/

Here's a little script that downloads and preprocessed the dataset for you: https://github.com/giulionf/download-camvid Note: It does simply download the videos and processes the frames, it does not split into train/test/val yet.