i have one small question about that the backbone from the file "Base-FSOD-C4.yaml" is set empty?

[SamXiaosheng]

No description provided.

[GuangxingHan]

The backbone has the default value as https://github.com/facebookresearch/detectron2/blob/main/detectron2/config/defaults.py#L133

[SamXiaosheng]

if not NAME，than its deault backbone is ResNET? 发自我的iPhone

…

------------------ Original ------------------ From: Guangxing Han ***@***.***> Date: Thu,Nov 24,2022 0:43 AM To: GuangxingHan/Meta-Faster-R-CNN ***@***.***> Cc: Sam ***@***.***>, Author ***@***.***> Subject: Re: [GuangxingHan/Meta-Faster-R-CNN] i have one small question aboutthat the backbone from the file "Base-FSOD-C4.yaml" is set empty? (Issue #19) The backbone has the default value as https://github.com/facebookresearch/detectron2/blob/main/detectron2/config/defaults.py#L133 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[GuangxingHan]

Yes, you are right.

[SamXiaosheng]

Hello, i have some confusion about your project, the second step "sh scripts/faster_rcnn_with_fpn_coco_base_classes_branch.sh"  and third step "scripts/few_shot_finetune_coco_resnet101.sh" have Superlative relationship? 

…

------------------ 原始邮件 ------------------ 发件人: "Guangxing ***@***.***>; 发送时间: 2022年11月24日(星期四) 上午7:54 收件人: ***@***.***>; 抄送: "生second ***@***.***>; ***@***.***>; 主题: Re: [GuangxingHan/Meta-Faster-R-CNN] i have one small question about that the backbone from the file "Base-FSOD-C4.yaml" is set empty? (Issue #19) Yes, you are right. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[GuangxingHan]

The second step is to train a separate Faster R-CNN detection head for base classes. The third step is to fine-tune the other detection head for few-shot novel classes.

[SamXiaosheng]

Hi, i found the second step use build_resnet_fpn_backbone and generated the final model with FPN module, but not used for third step. The third step use the pretraind model from the first step multi-stage3, edited in the file "xshot_finetune_coco_resnet101.yaml". at the same time, the RPN and ROI-Head are also different between the second and third step, so i doubt the difference between the second and third step in your exciting project.could you show the difference the second and third step in detail. i look forward to your reply in your convenience time. thanks in advance!

…

------------------ 原始邮件 ------------------ 发件人: "Guangxing ***@***.***>; 发送时间: 2022年11月25日(星期五) 凌晨2:09 收件人: ***@***.***>; 抄送: "生second ***@***.***>; ***@***.***>; 主题: Re: [GuangxingHan/Meta-Faster-R-CNN] i have one small question about that the backbone from the file "Base-FSOD-C4.yaml" is set empty? (Issue #19) The second step is to train a separate Faster R-CNN detection head for base classes. The third step is to fine-tune the other detection head for few-shot novel classes. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[GuangxingHan]

As shown in figure 2 of our paper, the second step and third share the same pre-trained backbone from the first step. The second step additionally learns the FPN head and detection head for base classes. The third step fine-tunes the Meta detection head pre-trained from the the first step. The idea is that we use two separate detection heads for base and novel classes respectively.