search

twangnh / Distilling-Object-Detectors

Implementation of CVPR 2019 paper: Distilling Object Detectors with Fine-grained Feature Imitation
MIT License
413 stars 70 forks source link

SSD #2

Open p517332051 opened 5 years ago

p517332051 commented 5 years ago

您好，请问有SSD的提升效果的表格吗

twangnh commented 5 years ago

Hi, currently I dont have ssd results

p517332051 commented 5 years ago

你好，我想问下那个resnet50-h是通道数为一半的resnet50而不是float16的resnet50？

---Original--- From: "twang"notifications@github.com Date: Fri, Jun 14, 2019 18:11 PM To: "twangnh/Distilling-Object-Detectors"Distilling-Object-Detectors@noreply.github.com; Cc: "Author"author@noreply.github.com;"疯狂三角肌"517332051@qq.com; Subject: Re: [twangnh/Distilling-Object-Detectors] SSD (#2)

Hi, currently I dont have ssd results

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

twangnh commented 5 years ago

The resnet50-h is obtained with halving each layer's channel number of resnet50

p517332051 commented 5 years ago

谢谢，还有一个问题您的论文中frcnn使用的是roipooling还是roialign？因为我在做过实验室roipooling和roipooling差别很大的。

---Original--- From: "twang"notifications@github.com Date: Wed, Jun 19, 2019 11:01 AM To: "twangnh/Distilling-Object-Detectors"Distilling-Object-Detectors@noreply.github.com; Cc: "Author"author@noreply.github.com;"疯狂三角肌"517332051@qq.com; Subject: Re: [twangnh/Distilling-Object-Detectors] SSD (#2)

the resnet50-h is obtained with halving each layer's channel number of resnet50

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

twangnh commented 5 years ago

roialign is used in the experiments

twangnh commented 5 years ago

I have seen your project on github, it is interesting that roialign and roipooling have so large difference on distillation results, do you have any update for the issue?

p517332051 commented 5 years ago

你可以用gluoncv去做下实验，差别很大。

---Original--- From: "twang"notifications@github.com Date: Tue, Jun 25, 2019 12:49 PM To: "twangnh/Distilling-Object-Detectors"Distilling-Object-Detectors@noreply.github.com; Cc: "Author"author@noreply.github.com;"疯狂三角肌"517332051@qq.com; Subject: Re: [twangnh/Distilling-Object-Detectors] SSD (#2)

I have seen your project on github, it is interesting that roialign and roipooling have so large difference on distillation results, do you have any update for the issue?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

UcanSee commented 4 years ago

你可以用gluoncv去做下实验，差别很大。 …

我想问一下您复现的Fine Grained feature imitation实验中roi Align和roi pooling效果会差多少？论文中roi Align提升约2.9%，那么如果是roi pooling大约提升多少呢？

p517332051 commented 4 years ago

看他github的

---Original--- From: "DliHjiang"<notifications@github.com> Date: Wed, Mar 18, 2020 14:51 PM To: "twangnh/Distilling-Object-Detectors"<Distilling-Object-Detectors@noreply.github.com>; Cc: "Author"<author@noreply.github.com>;"疯狂三角肌"<517332051@qq.com>; Subject: Re: [twangnh/Distilling-Object-Detectors] SSD (#2)

你可以用gluoncv去做下实验，差别很大。 …

我想问一下您复现的Fine Grained feature imitation实验中roi Align和roi pooling效果会差多少？论文中roi Align提升约2.9%，那么如果是roi pooling大约提升多少呢？

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

UcanSee commented 4 years ago

看他github的 … ---Original--- From: "DliHjiang"<notifications@github.com> Date: Wed, Mar 18, 2020 14:51 PM To: "twangnh/Distilling-Object-Detectors"<Distilling-Object-Detectors@noreply.github.com>; Cc: "Author"<author@noreply.github.com>;"疯狂三角肌"<517332051@qq.com>; Subject: Re: [twangnh/Distilling-Object-Detectors] SSD (#2)

你可以用gluoncv去做下实验，差别很大。 …

我想问一下您复现的Fine Grained feature imitation实验中roi Align和roi pooling效果会差多少？论文中roi Align提升约2.9%，那么如果是roi pooling大约提升多少呢？

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

哪里可以看到数据对比呢，我看您的GAN-KD论文里也没有Fine Grained feature imitation方法的roi align和roi pooling的对比，我最近尝试了带fpn的faster rcnn应用Fine Grained feature imitation方法，用的roi pooling，但是发现蒸馏后的结果比学生的基线还低，所以比较困惑，请问您那边实验的结果是什么样的呢？roi pooling会有提升吗？