No description, website, or topics provided.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
Fig1.jpg
Fig5.jpg
README.md
deploy.prototxt
solve.py
solver.prototxt
test.ipynb
train_val.prototxt

README.md

LSN

This code is for the paper "Linear Span Network for Object Skeleton Detection".

LSN builds on Holistically-Nested Edge Detection (HED) [1] with Liner Span Unit (LSU) under the guidance of linear span theory. LSU is used in three phases, feature linear span, resolution alignment, subspace linear span. A comparison of HED and LSN in principle is shown below. Each ellipse indicates a subspace spanned by feature maps from a convolutional layer. The union sets (the pink areas) imply the sum of the spanned subspaces. It can be seen that LSN principally spans a larger output space than HED does.

The comparision of the skeleton detection results of HED, SRN[2] and LSN on SK-LARGE dataset are shown below. From the results, it's easily to find that LSN explore more complementary features for skeleton detection and integrate them in a more effective way.

Installing

  1. Install prerequisites for Caffe (http://caffe.berkeleyvision.org/installation.html#prequequisites).
  2. Build HED (https://github.com/s9xie/hed). Supposing the root directory of HED is $HED.
  3. Copy files to $HED/example/LSN.

Training

  1. Download the (SK-LARGE) to $HED/data/, and do data augmentation follow instructions there;
  2. Download the Pre-trained VGG [3] model (VGG19). Copy it to $HED/example/LSN/
  3. Change the dataset path in $HED/example/LSN/train_val.prototxt
  4. Run solve.py in shell (or you could use IDE like Eclipse)
cd $HED/example/LSN/
python solver.py

Testing

  1. Change the dataset path in $HED/example/LSN/test.ipython.
  2. run test.ipython.

Evaluation

We use the evaluation code of [3] to draw the PR curve. The code can be download spb-mil.

NOTE: Before evaluation, the NMS is utilized. We use the NMS code in Piotr's edges-master.

Ref

[1] S. Xie and Z. Tu. Holistically-nested edge detection. In International Conference on Computer Vision, 2015

[2] Ke, W.,Chen, J.,Jiao, J., Zhao, G., Ye, Q.: SRN: side-out residual network for object symmetry detection in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition. (2017) 302-310

[3] S. Tsogkas and I. Kokkinos. Learning-based symmetry detection in natural images. In European Conference on Computer Vision