Non-person objects in COCO pretraining

[nemtiax]

When doing COCO pretraining, do you train on only person objects (only_person=True) or on all objects (only_person=False). It looks like the code as it stands does the latter, but I'm not sure if this is an indication of the intended algorithm, or if it's a relic of an ablation experiment. I had originally interpreted Table 4 from the paper as saying that only the person objects are used, but the code (

RAPiD/train.py

Line 45 in 12c9699

only_person = False if args.model == 'Hitachi80' else True

) has made me second-guess this.

[duanzhiihao]

or if it's a relic of an ablation experiment.

You have a good point, this line is indeed a relic of our ablation experiments.

However, we do train on only person objects, since args.model='rapid_pL1' by defualt, which makes only_person=True.

I had remove that line to clear up confusion.

[nemtiax]

Makes sense, thanks for the clarification!