NMS not working
ohlr opened this issue · 10 comments
I am using a network to generate more marks.
NMS, hierarchy thresh are at 0.5. As the image shows its not working. It works fine if I generate the marks with darknet directly.
I am on DarkMark 1.4.7
I can assure you it is working. What that tiny screenshot is showing me is you probably didn't train your network enough. What is max_batches set to, and how many classes do you have? Also see: https://www.ccoderun.ca/programming/2020-01-04_neural_network_training/#Details
Like I say, I use the same weights with
./darknet detector test mynet.data mynet.cfg mynet.weights -thresh 0.5 -dont_show -save_labels < train_new.txt
and it generates the labels fine. So it's not an issue of training.
Can you press the letter e in DarkMark to bring up the settings? It should look similar to this:
Note how all the predictions in my screenshot have just 1 box. I don't know why you would have multiple boxes, other than not enough training.
Did you build DarkHelp and DarkMark from source, or did you install the pre-built .deb files?
If you used the pre-built .deb files, can you follow the build steps instead and see if that fixes the problem? I'm wondering if maybe you and I are using different versions of Darknet and/or DarkHelp.
Building both from scratch, solved the issue. Thanks!
Just out of curiosity, what OS are you using, and how old is your version of Darknet?
I am on Linux Ubuntu
Darknet is at 1e3a616
Darknet commit hash 1e3a616 is from just a few days ago. This is the same version I'm using.
What version of Ubuntu? Can you paste the output of cat /etc/lsb-release?
Maybe another important thing I'll need to consider is what version NVIDIA drivers people are using. It may be simpler to not publish pre-built .deb files and have everyone build it themselves to avoid this issue in the future.
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=20.04
DISTRIB_CODENAME=focal
DISTRIB_DESCRIPTION="Ubuntu 20.04.2 LTS"
I don't have a GPU installed in the laptop. The only flag I set is OPENMP=1