search

ozendelait / rvc_devkit

Robust Vision Challenge Devkits
http://www.robustvision.net/
MIT License
107 stars 13 forks source link

Define joint obj. det. target format #7

Closed ozendelait closed 4 years ago

ozendelait commented 4 years ago

Potential candidates: ) COCO format ) OID format ) MaskRCNN/detectron2 format ) More?? As we use COCO-style format for the other tasks, I suggest to use it for boxable GT as well

michaelisc commented 4 years ago

I'd vote for COCO too to make sure things are consistent.

akuznetso commented 4 years ago

Sounds good, but image-level labels should be included in the annotations for OID - annotations are not exhaustive on image-level label level, i.e. if there is no positive or negative label 'Cat' on an image, there still might be an instance of 'Cat' that will not have a box.

Similarly, for instance segmentation data boxes and image-level labels should be included into annotations.

ozendelait commented 4 years ago

which datasets from RVC's obj. det. task support image level labels? COCO, OID, Obj365, MVS?

akuznetso commented 4 years ago

OID, I am not sure about the rest

rodrigob commented 4 years ago

Echoing @akuznetso comment, encoding the negative image level labels is particularly important.

wilson891226 commented 4 years ago

which datasets from RVC's obj. det. task support image level labels? COCO, OID, Obj365, MVS?

Obj365 can support image-level multi-labels.

tylin commented 4 years ago

COCO format does not have image-level label but seems like an easy addition to add.

ozendelait commented 4 years ago

@tylin / @wilson891226 : Are your datasets fully exhaustively annotated? Otherwise I don't understand where you are getting the information for the negative labels from?

wilson891226 commented 4 years ago

@tylin / @wilson891226 : Are your datasets fully exhaustively annotated? Otherwise I don't understand where you are getting the information for the negative labels from?

365 is not exhaustively annotated, but it can supply multi labels (i.e., the tags of the corresponding bounding boxes) associated with an image.

tylin commented 4 years ago

COCO exhaustively labels images with instance annotations (box/mask per object instance) and crowd annotations (segmentation mask covering multiple objects). For images with 10+ objects, 10 of the objects are labeled with instance annotations and a segmentation mask is used to cover the rest of objects. During inference, the predictions overlap with the crowd annotations are ignored.

ozendelait commented 4 years ago

Ok, maybe I don't see it, but what's the point of having those positive labels without the negative ones? Its trivial to create the prositive list on the fly in your framework if you need it from the boxable annotations...

ozendelait commented 4 years ago

Additional image labels may be added in the future; issue closed for now