images

cvzoya · Aug 4, 2017 · f7bee31 · f7bee31
1 parent ecbf703
commit f7bee31
Show file tree

Hide file tree

Showing 6 changed files with 5 additions and 0 deletions.
diff --git a/.DS_Store b/.DS_Store
diff --git a/data/.DS_Store b/data/.DS_Store
diff --git a/data/README.md b/data/README.md
@@ -42,6 +42,8 @@ Validation data (202 files):
 
 Note: the file paths in the training files point to this directory, so you can download the data files to this location in your cloned repo.
 
+Sample images and importance maps from the Massvis dataset:
+
 <img src="images/economist_daily_chart_103.png" width="21%" align="left"/>
 <img src="images/economist_daily_chart_103_imp.png" width="21%" align="left"/>
 <img src="images/economist_daily_chart_202.png" width="21%" align="left"/>

diff --git a/models/.DS_Store b/models/.DS_Store
diff --git a/models/README.md b/models/README.md
@@ -8,6 +8,9 @@
 The FCN-32s model was initialized from the pre-trained [VOC-FCN32s](https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn32s/caffemodel-url) model.
 The FCN-16s model was then initialized from the FCN-32s model. As specified in [fcn16/net.py](https://github.com/cvzoya/visimportance/blob/master/gdi/fcn16/net.py) it has an extra skip connection that was fine-tuned (all layers except score_sal and score_pool4 were frozen). This is the final model that we use for predicting graphic design importance.
 
+Here is a depiction of what this architecture with a skip connection roughly looks like (the darker layers are fixed, the lighter layers are fine-tuned on the importance data):
+<img src="sample_architecture.png" height="200">
+
 ## Massvis (Visualization)
 
   * FCN-32s: [massvis_fcn32.caffemodel](http://visimportance.mit.edu/data/massvis/massvis_fcn32.caffemodel)

diff --git a/models/sample_architecture.png b/models/sample_architecture.png