How to store yolo predictions into seperate files during testing ?

[BornInWater]

During inference on yolov3, i am currently parsing the bounding box co-ordinates displayed on the terminal by storing the entire thing into one file for all the images. This is very slow as i can process only one image at a time and hence takes a lot of time. Is there a way where i could test on multiple images at a time[in a batch] while being able to store predictions into different [mat/pkl] files?

[AlexeyAB]

Is there a way where i could test on multiple images at a time[in a batch]

You should implement it by yourself: https://github.com/AlexeyAB/darknet/issues/878#issuecomment-391926257

while being able to store predictions into different [mat/pkl] files?

It isn't implemented.

[BornInWater]

An unrelated question, after how many iterations is testing on validation done during the training process ? If thats configurable, where do i need to look for that? Also, is the validation accuracy at different snapshots stored in some file so as to pick the best snapshot?

Thank you for the help.

[AlexeyAB]

It isn't done automatically yet. You should do ./darknet detector map... manually for each 2000 iterations after ~2000 x classes iterations

[BornInWater]

Which code amongst those in "src" directory should i change in order to do this automatically?

[ritwikdubey]

while being able to store predictions into different [mat/pkl] files?

I intend to achieve something similar within the scope of darknet code.

Say, I have two set of weights, one for character segmentation and other for character recognition. Character segmentation returns following image (disregard the incorrect label names). I want to extract the image in the bounding box and run character recognition on it.

One way could be to write a wrapper script which does something like

Perform character segmentation using darknet.exe with flag -ext_output
extract sections from `predictions.png` based on output from darknet.exe
for each extracted image:
    run darknet.exe with character recognition weights
end

But I wish there's a way to perform bounding box extraction in darknet which would be more robust than my hacked up code.

[AlexeyAB]

@ritwikdubey Why you don't want to Detect characters as different objects (i.e. do Detection and Classification by using one Yolo v3 network)?

[ritwikdubey]

@AlexeyAB, interesting idea. I get a feeling that I might lose the sequence information of the characters. If A gets detected before K then the sequence might be read as AK.

Guess I can answer it myself, I could get the coordinates of center of each bounding box and sort them by their x-coordinates. This approach might be quite faster in detection as well, one call to detector will get me the characters by this approach compared to (num_of_characters+1) call to detector in my earlier idea.

Let me see how I can get it to work.