Is it necessary generate anchors for my own dataset to improve quality of the model or won't change with new anchors?
Yes it is necessary. Yo can have a look here an try to follow this tutorial:
https://github.com/Jumabek/darknet_scripts
@Savash2016 Did you try to do it by yourself? By this script
I think that script is only for VOC.
Yes, it is. I've made it more general for generating anchors for my dataset. But when I put new anchors in cgf file, results became worse: bounding boxes were in the right place, but they were very small
so what do you think we should do? Just leave the original anchors as is?
@borasy If there are no others' experience of using their own dataset with their own anchors. These anchors works well.
alright.. i'll leave it as is then.. Thanks for your experiment. But later if you gain better results with your own anchors please let me know.
@Savash2016 @dimaxano i have a custom dataset for 6 labels is it good to generate anchors ??or by using default anchors how will the result vary any theoretical references
I would propose to try with script for generating custom anchors for your dataset. Theoretically this is a right way. But I didn't manage to get better results with custom anchors, as i think, cause of wrong use of script for anchor generating)
Summary: try custom anchors, but default are also OK
I've written a small script so that you can convert to voc labels and then generate anchors boxes based on the above code. Let me know if there is any issues
https://github.com/ttorkar/Voc_Label
@dimaxano have you changed the object_scale parameter in your cfg file? the bounding boxes generated by the script have a scale of 0 to 1
Most helpful comment
I've written a small script so that you can convert to voc labels and then generate anchors boxes based on the above code. Let me know if there is any issues
https://github.com/ttorkar/Voc_Label