Training FCN-32S on the NYUD dataset

[Feng Gu]

I have been experimenting with FCN-16S and FCN-8S on the NYUD dataset. I downloaded the code from https://github.com/shelhamer/fcn.berkeleyvision.org, and modified the scripts pascalcontext-fcn8s and pascalcontext-fcn16s with respect to the training (colour only) on the NYUD dataset. However, I got really poor results. So then I use the provided scripts in nyud-fcn32s-color to train a FCN-32S model, the intermediate evaluation gives the following results:

 I0622 12:39:10.599292  3302 sgd_solver.cpp:106] Iteration 151980, lr = 1e-10

>>> 2016-06-22 12:39:18.745470 Begin seg tests

>>> 2016-06-22 12:41:27.654631 Iteration 152000 loss 695537.10345

>>> 2016-06-22 12:41:27.654709 Iteration 152000 overall accuracy 0.21258416853

>>> 2016-06-22 12:41:27.654743 Iteration 152000 mean accuracy 0.025

>>> 2016-06-22 12:41:27.654872 Iteration 152000 mean IU 0.00531460421324

>>> 2016-06-22 12:41:27.654947 Iteration 152000 fwavacc 0.0451920287095

The paper reported much better performance. In addition, the loss doesn't seem to decrease as quickly as expected. BTW I'm using the version of dataset provided here: https://github.com/s-gupta/rcnn-depth for training. Could you someone please tell me what could possibly went wrong in my experiments.

[ustc...@gmail.com]

Hi, I did the same thing as you and I met the same prooblem. Did you sovle it? By the way I find the loss is 1e5 level because the 'normalize' in the loss layer is set to false. 

在 2016年6月22日星期三 UTC+8下午8:33:38，Feng Gu写道：