search

scaelles / DEXTR-PyTorch

Deep Extreme Cut http://www.vision.ee.ethz.ch/~cvlsegmentation/dextr
GNU General Public License v3.0
843 stars 153 forks source link

Performance of DEXTR with 2 or 3 points #32

Closed SebastienLinker closed 4 years ago

SebastienLinker commented 4 years ago

Hi,

DEXTR is an amazing tool. Out of curiosity, have you tried training DEXTR with only 2 or 3 extreme points? I am interested in using this tool to segment long and thin objects like tree trunks, springs, road lanes ...

Thanks

kmaninis commented 4 years ago

Hello, thanks for the kind words. With 2 or 3 extreme points we are not able to recover the bounding box (in most cases). The bounding box leads to substantial performance improvements (Table 2, first row in the paper).

scaelles commented 4 years ago

As @kmaninis mentioned it is quite important for our method to use a bounding boxes. If you think you can obtain one from the 2 or 3 extreme points by assuming some parameters ie clicking the top and bottom and assuming always a fix width, then you could give it try!

SebastienLinker commented 4 years ago

Thanks

It actually works on those long and thin objects, where 2 extreme points are physically close to each others