search

zhenyuw16 / noisyboundaries

Apache License 2.0
40 stars 6 forks source link

The division of coco data is different from semi-supervised detection methods on COCO? #3

Open Sunting78 opened 2 years ago

Sunting78 commented 2 years ago

1) There are 5 folds in semi-supervised detection methods on COCO. Do you the same ? 2) The paper shows: For the 30% setting,we simply use the 35k subset of the COCO 2014 validationset as labeled images. But，in you code 'prepare_coco_data.py' , there is only used instances_train2017.json. https://github.com/zhenyuw16/noisyboundaries/blob/0c83823fe3634edb1274f4b20829822e49a25d65/scripts/coco/prepare_coco_data.py#L35

Look forward to your reply. Thanks.

zhenyuw16 commented 2 years ago

For 1% to 10%, we use the prepare_coco_data.py to generate data. Besides this setting, we also use the 35k subset of the COCO 2014 validation set as labeled images. This set the widely used COCO 2014 valminusminival. The 35k is about 30% of the COCO 2017 so we simply denote it as 30%. Since COCO 2014 valminusminival is a predefined set, you do not need to generate it with this script.

Sunting78 commented 2 years ago

Thanks. But i want to know, is this the data division firstly proposed in this paper? Why this data division diffierent from semi-supervised detection for comparing the results of Boxes AP?

zhenyuw16 commented 2 years ago

As I know, this dividion is firstly proposed in "Data Distillation: Towards Omni-Supervised Learning".

Sunting78 commented 2 years ago

As I know, this dividion is firstly proposed in "Data Distillation: Towards Omni-Supervised Learning".

Got it. Thanks.