Automatically accept all darknet predictions?

[bcherb2]

I see there is a keyboard shortcut to accept all predictions within an image, however I do not see a way to accept all predictions for a project? I have trained an initial model off of a couple hundred images, and would then like to apply that to several thousand images. Afterwards I would then go through and verify the tags are correct, but automatically tagging each frame seems like it would save quite a bit of time. Considered this feature?

This is amazing software by the way! Works well over ssh -X, too

[stephanecharette]

Thanks. Unfortunately not in DarkMark, I don't have an option for that.

Next easiest thing I can think of is to use DarkHelp Server with the option to create the .txt files for each image. Then you let it process your thousands of images. DarkMark will then automatically import those .txt files. Not straightforward, but may be worthwhile if you have so many images.

If you want to try, DarkHelp Server is documented here: https://www.ccoderun.ca/darkhelp/api/Server.html Note you already have it installed since DarkMark uses DarkHelp. But you would have to configure it to run the way you need.

[stephanecharette]

If you do try, the option you want is the one called darkhelp/server/settings/save_txt_annotations.

[bcherb2]

Thanks, I will look into that. I tried the pseudo-labeling CLI for darknet, and I got unexpected results. Both
./darknet detector test ..... -thresh 0.25 -dont_show -save_labels > results.txt and
./darknet detector test ..... -thresh 0.25 -dont_show -ext_output < train.txt > results.txt

both result in many entries like this:

/home/ben/VisionProjects/Palletsv3/darkmark_image_cache/zoom/00001739.jpg: Predicted in 11.293000 milli-seconds. Person: 99% Person: 99% Person: 99% Enter Image Path: Detection layer: 139 - type = 28 Detection layer: 150 - type = 28 Detection layer: 161 - type = 28

Upon further inspection, it looks like it also wrote *.txt files next to the unmarked images, and refreshing and rescanning in DarkMark was able to find the new marks.

[stephanecharette]

The problem with using Darknet is I believe that it will generate some annotations with negative coordinates or widths/heights that extend beyond the edges of the image. May want to have a browse through the .txt files to see.

However, I think DarkMark fixes those up when it imports them. That was so long ago I don't recall anymore, you'll have to test it and see.