I try to train FCN by myself, but I failed to obtain the "PASCAL-context 59-class" dataset

[Yanye Li]

Recently I try to train a FCN by myself following this example https://gist.github.com/shelhamer/80667189b218ad570e82

In the example, it says it trained for the PASCAL-context 59-class (60 including background) task, with the per-channel mean

B 104.00698793 G 116.66876762 R 122.67891434

I am not sure whether PASCAL-context means VOC2010.

Besides, if it meant VOC2010, I cannot find the 59-class segmentation as ground truth labels.

[Evan Shelhamer]

This is the PASCAL-Context data set: http://www.cs.stanford.edu/~roozbeh/pascal-context/

To reproduce the results it is important to

- fine-tune from the VGG16 classifier, which is included in the model zoo

- remember to initialize the interpolation weights, as done by solve.py

- use batch size one so that the net can be reshaped for each input (we don't crop/scale/warp images—we take them as they are—and note that this reshaping is essentially free)

We'll post new examples soon now that the Crop layer and coordinate mapping through Python net spec are in master.

Evan Shelhamer

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/5c5ca576-6c9b-424e-9dfa-7908107e6241%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Yanye Li]

Thanks for answering my questions!

I went http://www.cs.stanford.edu/~roozbeh/pascal-context/ and downloaded the file trainval.tar.gz, which contains ground truth labels rangeing 1 to 459 rather than 1 to 59.

What should I do to prepare the input data for the FCN input? Should I reduce the labels?

Besides, I noticed in train_val.prototxt of FCN, the data and gt-label are separately imported from two LMDB, 'pascal-context-train-lmdb' and 'pascal-context-train-gt59-lmdb'. However, in other net, e.g. LeNet on MNIST, MNIST data and label are imported from just one LMDB. Is this because in FCN, the ground truth labels are actually label-maps rather than a single label?

在 2016年4月12日星期二 UTC+8下午1:48:44，Evan Shelhamer写道：

[Evan Shelhamer]

​ground truth labels rangeing 1 to 459 rather than 1 to 59.
What should I do to prepare the input data for the FCN input?

The PASCAL-Context data includes a lot of (noisy) labels in its raw form​. For this reason its authors identified a 59 class subset with clearer annotations and these are the classes we used. I'll try to post the exact mapping soon.

Until then, please check out the master edition of our reference FCNs and related code: http://fcn.berkeleyvision.org. The data layer in particular illustrates how to handle the image and ground truth, although for PASCAL-Context you will still need to handle the label mapping.

Evan Shelhamer

To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/5c3f8a0b-b316-44c7-bb73-0b23d18bc603%40googlegroups.com.