Repository navigation
Fast version #1
Description
Dude, the prec/recall function is very slow. Use numpy if possible for batch calculations:
def calc_precision_recall(contours_a, contours_b, threshold):
x = contours_a
y = contours_b
t1 = np.tile(x, (len(y),1))
t2 = np.repeat(y, len(x), axis=0)
t = np.concatenate( (t1,t2), axis=1)
dx = (t[...,0] - t[...,2]) * (t[...,0] - t[...,2])
dy = (t[...,1] - t[...,3]) * (t[...,1] - t[...,3])
thr = (dx + dy) < (threshold * threshold)
blocks = np.split(thr, len(y))
block_has_hit = [np.any(x == True) for x in blocks]
top_count = np.sum(block_has_hit)
precision_recall = top_count/len(y)
return precision_recall, top_count, len(y)
It should be the same, you can validate it with your own data.
Thanks, Balint
Reacted by Matiur Rahman Minar
Activity
bpfliegel commentedon Oct 21, 2019
Even faster to keep the outer loop:
Thanks, all the best!
minar09 commentedon Oct 22, 2019
Thank you very much, @bpfliegel for your suggestions. Appreciate it.
bpfliegel commentedon Oct 22, 2019
Thanks to you for your work! :)
minar09 commentedon Apr 18, 2020
Hey, @bpfliegel , just wondering could you please create a pull request with your improvements? That would be really great. Thank you very much.
bpfliegel commentedon Apr 19, 2020
@minar09 too busy, sorry, but I posted you the source code above ^^^ thanks
minar09 commentedon Apr 20, 2020
@bpfliegel , same here. No problem. Thank you very much for the codes. I will try to update the repository.
theodumont commentedon Nov 17, 2020
Hi! Thanks for the repo.
I implemented @bpfliegel's 2nd version and did a PR, all the details are there: #2
minar09 commentedon Nov 17, 2020
Thank you @theodumont for your contributions.