search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.39k stars 652 forks source link

Blendmask training on custom dataset but category ID or class_name wrong #562

Open kimile599 opened 2 years ago

kimile599 commented 2 years ago

I was able to train the blendmask and inference the model on my custom dataset which is prepared in coco format. The segmentation was good but the class_name on the bounding box was still person and bicycle, which obviously came from the coco2017_train dataset. I can see my dataset registered correct and the thing_class also seemed to be right.

Have no idea how to solve it.

Stenyears commented 2 years ago

i have the same question .did you solve it?

rasheedtobi commented 2 years ago

In AdelaiDet/configs/BlendMask/Base-BlendMask.yaml, the default settings is: DATASETS: TRAIN: ("coco_2017_train",) TEST: ("coco_2017_val",)

which references to the classes in coco dataset. First of all, change the default and use the registered name

secondly, have to define the classes and register the dataset in demo.py file.