for tack1, we are required to submit the matched results.
is there any requirement for the matching algorithm. e.g.
-
only KNN retrieval (like track 2) is allowed
-
if we assume one-to-one match (i.e. one reference is only match to one query), there are many post-processing tricks we can use for bipartite match. Are such assumptions and methods allowed?
-
model the matcher as another deep network
-
assume match is random, i.e. if we randomly divide the reference images into groups (buckets) of m images each, number of matches found for each group is fairly constant
-
"There are a total of 50K query images, of which 10K have matches in the reference set,
while the remaining 40K do not. " Use this prior knowledge in our modeling. But do stage2 images follow the same rule? -
train classifier to identify disractor images?
-
use meata data like image size, jpg header, etc
-
check for jpg recompression artifacts or other image forensics methods,