search

yanxp / MetaR-CNN

Meta R-CNN : Towards General Solver for Instance-level Low-shot Learning
https://yanxp.github.io/metarcnn.html
177 stars 23 forks source link

Confusion about coco split #8

Closed lzhnb closed 4 years ago

lzhnb commented 4 years ago

In your paper, the tasks contain MS COCO to PASCAL.

And i'm curious about the classes-split of this task:

  1. Choosed classes from voc (20 classes) as novel classes and let rest classes(60 classes) be the bases classes.

    This is my first comprehension of this task after reading paper

  2. Choosed classes from your split_file (three txt - 5 classes) as novel classes and the rest classes in voc classes (15 classes) as base classes.

    I've change your train example and train successfully in coco dataset, i find the base and novel classes also depends on the meta_type which splits the voc_classes into 15 base classes and 5 novel classes

Thanks

yanxp commented 4 years ago

Yes, It 's right.

lzhnb commented 4 years ago

So, the second comprehension is right?

zb1439 commented 4 years ago

Samely curious about this issue! I'm not clear with class split with COCO dataset (is it 60-20 split as previous work on few shot detection like LSTD or using VOC classes only with 15-5 split) ? In addition, could you please release your interface for COCO dataset training? Thx and looking forward to reply.