Curve Fragment Ground-Truth Dataset (CFGD)

                                     Curve Fragment Ground-Truth Dataset

We develop a new type of human annotated ground truth centered on contour fragment. The goal of this collection is not distinguished objects but curves evident in image data. This contribute refer to On Evaluating Methods for Recovering Image Curve Fragments. I also put my presentation in workshop of CPVR 2012 here.

We prepared two versions of this dataset. One version is in fine scale, which is collected from human objects using zoomed in images. The other version is in coarse scale, which is collected from human objects with original size images. Each version of our dataset is presented in both binary maps and curve fragments files. Binary maps can be used for evaluating edge detection or boundary detection methods. Curve fragments files can be used to evaluation curve fragment extracting methods, which is also detailed in that paper. Curve Fragment Ground-Truth Dataset (CFGD) is available.

Matlab version of Curve Fragment Evaluation Code is available If you would like to evaluation your bottom up boundary detection algorithm using Berkeley's Evaluation Method, but on our non-semantic or low-semantic ground truth. We provide the Boundary detection Evaluation code which is implemented with the same algorithm as BSDS but replace the our ground truth.

Because of the clean up of the evaluation code and different setting of varying threshold, Precision-Recall curve may be different from what is showed in the paper. However, the comparative performance between different algorithms are still the same.

This webpage is currently maintained by Yuliang Guo (Yuliang_Guo at Brown.edu). Feel free to contact me if you have problems using our dataset or algorithms.