Giters

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.

Home Page:https://git.io/AdelaiDet

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Fully-supervised training of finetuned BoxInst model on custom dataset

ameyparanjape opened this issue · comments

commented

I have an experiment where my train data is as follows:

Format - coco style custom dataset with 6 classes
(a) 20% of the train data has mask+bbox annotations
(b) Remaining 80% of the train data has only bbox annotations.

I was able to train/finetune a BoxInst model (weakly supervised instance segmentation) using (b), now I want to further finetune the model using (b) (utilizing the mask annotations) as a fully supervised training. BoxInst paper does talk about this, but I wasn't able to find the code changes for this.

Any help?

commented

I was able to figure this one out.
Changing config parameter

BOXINST:
    ENABLED: false

This will trigger training as a regular Instance Segmentation model without the BoxInst losses being computed and optimized.

commented

@ameyparanjape i have trained a custom model using detectron 2 custom tutorial to get segmentation masks from Boxinst training. However i get zero AP. can you guide me if i am missing something ? does train_net.py have some configuration that is not in detectron2 demo ?