How to zero-shot transfer to PASCAL VOC? The config file does not provide any information about PASCAL VOC

[LemonTwoL]

1. Please check default.yml, if we should add the node of test just like train and val nodes.
2. default.yml also did not provide the setting of pascal voc dataset and coco dataset

[xvjiarui]

Hi @XiaominLi1997 The config is here https://github.com/NVlabs/GroupViT/blob/main/configs/default.yml#L90

[Qi-Zhangyang]

I also want to know the evaluation part. Does the cls part only include the ImageNet? And the seg part includes Pascal VOC, Pascal Context and COCO? Thank you very much.

config file: evaluate: eval_only: false eval_freq: 1 task: - cls - seg cls: save_best: true template: subset seg: save_best: true cfg: segmentation/configs/base/datasets/pascal_voc12.py template: simple opts: []

[Qi-Zhangyang]

I also want to know the evaluation part. Does the cls part only include the ImageNet? And the seg part includes Pascal VOC, Pascal Context and COCO? Thank you very much.

config file: evaluate: eval_only: false eval_freq: 1 task: - cls - seg cls: save_best: true template: subset seg: save_best: true cfg: segmentation/configs/base/datasets/pascal_voc12.py template: simple opts: []

[LemonTwoL]

Hi @XiaominLi1997 The config is here https://github.com/NVlabs/GroupViT/blob/main/configs/default.yml#L90

Thanks. And I have another question, in the reference phase, can GroupViT only group the objects appearring in the training samples? If GroupViT can group objects unseen in the training phase.

Thank you again for your patient reply.

[xvjiarui]

Hi @XiaominLi1997 Since the training samples are noisy, it may cover a wide range of concepts. Some of them may overlap with our evaluation dataset, e.g. Pascal VOC. But we didn't explicitly evaluate the out of distribution grouping since it's hard to exclude some objects from noisy training set.

[xvjiarui]

I also want to know the evaluation part. Does the cls part only include the ImageNet? And the seg part includes Pascal VOC, Pascal Context and COCO? Thank you very much.

config file: evaluate: eval_only: false eval_freq: 1 task: - cls - seg cls: save_best: true template: subset seg: save_best: true cfg: segmentation/configs/base/datasets/pascal_voc12.py template: simple opts: []

The cls part is referring to the ImageNet zero-shot classification. You may change the evaluation dataset following instructions here. By default, we are using Pascal VOC dataset.