diff --git a/README.md b/README.md index e239529cfb..268f6daaa4 100644 --- a/README.md +++ b/README.md @@ -29,16 +29,19 @@ We provide the pretrained model and training/testing code for the edge detection 0. Download the pretrained model (56MB) from (http://vcl.ucsd.edu/hed/hed_pretrained_bsds.caffemodel) and place it in examples/hed/ folder. ### Installing + 0. Install prerequisites for Caffe(http://caffe.berkeleyvision.org/installation.html#prequequisites) 0. Modified-caffe for HED: https://github.com/s9xie/hed.git ### Training HED To reproduce our results on BSDS500 dataset: + 0. data: Download the augmented BSDS data (1.2GB) from (http://vcl.ucsd.edu/hed/HED-BSDS.tar) and extract it in data/ folder 0. initial model: Download fully convolutional VGG model (248MB) from (http://vcl.ucsd.edu/hed/5stage-vgg.caffemodel) and put it in examples/hed folder 0. run the python script **python solve.py** in examples/hed ### Testing HED + Please refer to the IPython Notebook in examples/hed/ to test a trained model. The fusion-output, and individual side-output from 5 scales will be produced after one forward pass. Note that if you want to evaluate the results on BSDS benchmarking dataset, you should do the standard non-maximum suppression (NMS) and edge thinning. We used Piotr's Structured Forest matlab toolbox available here **https://github.com/pdollar/edges**. Some helper functions are also provided in the [eval/ folder](https://github.com/s9xie/hed_release-deprecated/tree/master/examples/eval). @@ -50,7 +53,6 @@ Note that if you want to evaluate the results on BSDS benchmarking dataset, you ### Precomputed Results If you want to compare your method with HED and need the precomputed results, you can download them from (http://vcl.ucsd.edu/hed/eval_results.tar). - ### Acknowledgment: This code is based on Caffe. Thanks to the contributors of Caffe. Thanks @shelhamer and @longjon for providing fundamental implementations that enable fully convolutional training/testing in Caffe.