search

s9xie / hed

code for Holistically-Nested Edge Detection
Other
1.8k stars 530 forks source link

There is a black line in some train images, why do that? #55

Open idealwei opened 6 years ago

idealwei commented 6 years ago

like 100080.jpg under folder HED-BSDS/train/0.0_1_1/100080.jpg

LemonMi commented 6 years ago

I have the same problem with you.There is a strange offset for the output image when i use the pretrained model.

arthur-qiu commented 6 years ago

Same problem. I think there are some problems in the pretrained model. An easy solution is modify your input, test it, and then modify it again. Different lay has different padding. For example, My input is 512x512, and output is from out3 (around 35pixel black line on the top and left). To solve it, first, I modify my input as 547x547(add blank to the right and bottom of original input). Then use HED. Store the [35:547, 35:547] as result. I think it is one of best choices. More tips: in your original input, you'd better leave some empty space. That's due to HED will diffuse the original image mildly. For me, l resize the 512 to 496 and leave 8pixel around.

ghost commented 6 years ago

@here, its pretty annoying indeed, but you can meet the algorithm half way, it's fortunately just a translation, by ~33 pixels along the x and y axes. So running the output from the net like this net.blobs['sigmoid-dsn1'].data[0][0][33:][:,33:] should get rid of the border and give you your output array.

kzq666666 commented 2 years ago

Thanks for your advice, I meet the same problem. It does works.