pinakinathc
commented
3 years ago Thanks for this amazing work !!
@MarkMoHR can you release the sequences for these training/validation/test sketch-image pair sequences?
This will help me benchmark your paper more consistently.
Considering the objects presented in the sketches of SketchyScene
are not 100% aligned with the reference images (as explained in Sec.3), we se-
lected sketch-photo pairs whose semantic-IoU are higher than 0.5 (2,472 pairs
for training and validation while 252 for testing).
This is an extension of Issue #21 I was checking the correspondence of images and sketches. Here is just one example -- can we expect to have instance-level mapping? (i.e., for a more proper fine-grained retrieval task)