shelhamer / clockwork-fcn Goto Github PK

Hi, thank you for sharing this great work.
I want to training this network,but I facing some problem about training detail. If you release the train.prototxt and solver.prototxt, I will be very grateful to you.
Thank you !

I have tried the "pascal-translate-exp.ipynb", and find some typos,

1. in shell [2], "caffe.set_device(1)", I think it's better to use "caffe.set_device(0)".
2. in shell [5], '../nets/fcn-pool3-pascal.caffemodel', '../nets/fcn-pool4-pascal.caffemodel', but in your "clockwork-nets" package, the corresponding file is "voc-fcn-pool3.caffemodel" and "voc-fcn-pool4.caffemodel". I think it's better to use the same names.

I want to use it to infer a new movie . how to do it ?

When I was testing the code on cityscapes, I encountered an error : File "datasets\cityscapes.py", line 17, in init labels = import('labels') ImportError: No module named labels. How can I get this labels module? Thank you

Hi @shelhamer

How to use and test this code ?can you give me a guidance? Thanks

First for the installation it says to install everything in "requirements.txt" but file is missing.

Also, I am trying to test your code but I need the "prostage-cityscapes-fcn8s.prototxt". In general if I want to test all the prototxt are missing, can you provide them?

Thanks

nets/cityscapes-fcn8s.prototxt is lost

Hi, could you send me youtube objects segmentation dataset, I cannot find it ?
Thank you.

Hi, shelhamer. I have test your code and it works fine with cityscapes.
I'm using CamVid dataset and i want to use your network on this dataset.
Could you provide caffemodel for this dataset?
Thank you.

Hi, shelhamer. I have test your code and it works fine.
While I haven't find the scripts or steps about how to train this clockwork-fcn network. Would you give me a hand?
Thanks a lot.

In this paper, "Clockwork Convnets for Video Semantic Segmentation", The NYUDv2 dataset [6] collects short RGB-D clips and includes a segmentation benchmark with high-quality but temporally sparse pixel annotations (every tenth video frame is labeled). We run on video from the “raw” clips subsampled 10X and evaluate on every labeled frame. But i am confused how to get the every 10th labeled ground truth, maybe I have to annotate them by myself, because I have not found the download link in http://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.