search

xavysp / DexiNed

DexiNed: Dense EXtreme Inception Network for Edge Detection
MIT License
714 stars 155 forks source link

Could you publish NYUD preprocessing script? #20

Open shkarupa-alex opened 4 years ago

shkarupa-alex commented 4 years ago

I'm interested in how do you extract boundaries from NYUDv2

xavysp commented 4 years ago

I'm interested in how do you extract boundaries from NYUDv2

Hi @shkarupa-alex, the manuscript of DexiNed present just edges not boundaries, we are proposing to train just one time on BIPED dataset and evaluate it on the other edge-based datasets, of course, we have tested in a non-edge dataset like PASCAL, NYUD, or even BSDS, just to show the difference, and state that at least for edge detection we do not need train the model on each database. So we can use DexiNed and test any image from the internet (even NYUDv2) and test it. concerning boundary dataset, DexiNed presents the best perceptual appearance but qualitatively we cannot reach a state-of-the-art because the GT is from segmentation annotation and they has many biases.

Sorry, let me know if I answered your question. The image preprocessing (before predict with DexiNed) is the same for all input images.

Cheers

Xavier

claudia-new commented 3 years ago

I have the same issue, how should I preprocess the data if I want to test other datasets like BSDS? In both BIPED and MULTICUE, the test list contain images and corresponding edge_maps, for example in test_pair.lst

test/imgs/4100_left_2268.png test/edge-maps/4100_left_2268.6.png test/imgs/4101_left_0382.png test/edge-maps/4101_left_0382.6.png test/imgs/4108_left_0082.png test/edge-maps/4108_left_0082.6.png test/imgs/4113_left_0082.png test/edge-maps/4113_left_0082.6.png test/imgs/4114_left_0611.png test/edge-maps/4114_left_0611.6.png test/imgs/4176_left_0273.png test/edge-maps/4176_left_0273.6.png test/imgs/4178_left_0052.png test/edge-maps/4178_left_0052.6.png test/imgs/4179_left_0315.png test/edge-maps/4179_left_0315.6.png test/imgs/4187_left_0071.png test/edge-maps/4187_left_0071.6.png test/imgs/4209_left_0034.png test/edge-maps/4209_left_0034.6.png

but BSDS and NYU don't, how should I set the test list respectively? Or is the edge-maps necessary to test any data? If not, how can I set the parameters? @xavysp

xavysp commented 3 years ago

I have the same issue, how should I preprocess the data if I want to test other datasets like BSDS? In both BIPED and MULTICUE, the test list contain images and corresponding edge_maps, for example in test_pair.lst

test/imgs/4100_left_2268.png test/edge-maps/4100_left_2268.6.png test/imgs/4101_left_0382.png test/edge-maps/4101_left_0382.6.png test/imgs/4108_left_0082.png test/edge-maps/4108_left_0082.6.png test/imgs/4113_left_0082.png test/edge-maps/4113_left_0082.6.png test/imgs/4114_left_0611.png test/edge-maps/4114_left_0611.6.png test/imgs/4176_left_0273.png test/edge-maps/4176_left_0273.6.png test/imgs/4178_left_0052.png test/edge-maps/4178_left_0052.6.png test/imgs/4179_left_0315.png test/edge-maps/4179_left_0315.6.png test/imgs/4187_left_0071.png test/edge-maps/4187_left_0071.6.png test/imgs/4209_left_0034.png test/edge-maps/4209_left_0034.6.png

but BSDS and NYU don't, how should I set the test list respectively? Or is the edge-maps necessary to test any data? If not, how can I set the parameters? @xavysp

The easiest way is to put, in the data dir (the dir that has lenna image for single testing), the whole of images from NYUD or BSDS. As you can see in the CLASSIC dataset, the GT is not necessary in the testing stage.

Cheers,

Xavier

claudia-new commented 3 years ago

Thank you, I have solved the problem, it's about the dataset list, after I rewrite the test list and change the path, it's done. Thanks @xavysp