# Group Normalization for Mask R-CNN
<div align="center">
<img src="gn.jpg" width="700px" />
</div>
## Introduction
This file provides Mask R-CNN baseline results and models trained with [Group Normalization](https://arxiv.org/abs/1803.08494):
```
@article{GroupNorm2018,
title={Group Normalization},
author={Yuxin Wu and Kaiming He},
journal={arXiv:1803.08494},
year={2018}
}
```
**Note:** This code uses the GroupNorm op implemented in CUDA, included in the Caffe2 repo. When writing this document, Caffe2 is being merged into PyTorch, and the GroupNorm op is located [here](https://github.com/pytorch/pytorch/blob/master/caffe2/operators/group_norm_op.cu). Make sure your Caffe2 is up to date.
## Pretrained Models with GN
These models are trained in Caffe2 on the standard ImageNet-1k dataset, using GroupNorm with 32 groups (G=32).
- [R-50-GN.pkl](https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl): ResNet-50 with GN, 24.0\% top-1 error (center-crop).
- [R-101-GN.pkl](https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/47592356/R-101-GN.pkl): ResNet-101 with GN, 22.6\% top-1 error (center-crop).
## Results
### Baselines with BN
<table><tbody>
<!-- START E2E MASK RCNN BN TABLE -->
<!-- TABLE HEADER -->
<!-- Info: we use wrap text in <sup><sub></sub><sup> to make is small -->
<th valign="bottom"><sup><sub> case </sub></sup></th>
<th valign="bottom"><sup><sub>type</sub></sup></th>
<th valign="bottom"><sup><sub>lr<br/>schd</sub></sup></th>
<th valign="bottom"><sup><sub>im/<br/>gpu</sub></sup></th>
<th valign="bottom"><sup><sub>train<br/>mem<br/>(GB)</sub></sup></th>
<th valign="bottom"><sup><sub>train<br/>time<br/>(s/iter)</sub></sup></th>
<th valign="bottom"><sup><sub>train<br/>time<br/>total<br/>(hr)</sub></sup></th>
<th valign="bottom"><sup><sub>inference<br/>time<br/>(s/im)</sub></sup></th>
<th valign="bottom"><sup><sub>box<br/>AP</sub></sup></th>
<th valign="bottom"><sup><sub>mask<br/>AP</sub></sup></th>
<th valign="bottom"><sup><sub>model id</sub></sup></th>
<tr>
<td align="left"><sup><sub>R-50-FPN, BN*</sub></sup></td>
<td align="left"><sup><sub>Mask R-CNN</sub></sup></td>
<td align="left"><sup><sub>2x</sub></sup></td>
<td align="right"><sup><sub>2</sub></sup></td>
<td align="right"><sup><sub>8.6</sub></sup></td>
<td align="right"><sup><sub>0.897</sub></sup></td>
<td align="right"><sup><sub>44.9</sub></sup></td>
<td align="right"><sup><sub>0.099 + 0.018</sub></sup></td>
<td align="right"><sup><sub>38.6</sub></sup></td>
<td align="right"><sup><sub>34.5</sub></sup></td>
<td align="right"><sup><sub>35859007</sub></sup></td>
</tr>
<tr>
<td align="left"><sup><sub>R-101-FPN, BN*</sub></sup></td>
<td align="left"><sup><sub>Mask R-CNN</sub></sup></td>
<td align="left"><sup><sub>2x</sub></sup></td>
<td align="right"><sup><sub>2</sub></sup></td>
<td align="right"><sup><sub>10.2</sub></sup></td>
<td align="right"><sup><sub>0.993</sub></sup></td>
<td align="right"><sup><sub>49.7</sub></sup></td>
<td align="right"><sup><sub>0.126 + 0.017</sub></sup></td>
<td align="right"><sup><sub>40.9</sub></sup></td>
<td align="right"><sup><sub>36.4</sub></sup></td>
<td align="right"><sup><sub>35861858</sub></sup></td>
</tr>
<!-- END E2E MASK RCNN BN TABLE -->
</tbody></table>
**Notes:**
- This table is copied from [Detectron Model Zoo](https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md#end-to-end-faster--mask-r-cnn-baselines).
- BN<sup>*</sup> means that BatchNorm (BN) is used for pre-training and is frozen and turned into a per-channel linear layer when fine-tuning. This is the default of Faster/Mask R-CNN and Detectron.
### Mask R-CNN with GN
#### Standard Mask R-CNN recipe
<table><tbody>
<!-- START E2E MASK RCNN GN TABLE -->
<!-- TABLE HEADER -->
<!-- Info: we use wrap text in <sup><sub></sub><sup> to make is small -->
<th valign="bottom"><sup><sub> case </sub></sup></th>
<th valign="bottom"><sup><sub>type</sub></sup></th>
<th valign="bottom"><sup><sub>lr<br/>schd</sub></sup></th>
<th valign="bottom"><sup><sub>im/<br/>gpu</sub></sup></th>
<th valign="bottom"><sup><sub>train<br/>mem<br/>(GB)</sub></sup></th>
<th valign="bottom"><sup><sub>train<br/>time<br/>(s/iter)</sub></sup></th>
<th valign="bottom"><sup><sub>train<br/>time<br/>total<br/>(hr)</sub></sup></th>
<th valign="bottom"><sup><sub>inference<br/>time<br/>(s/im)</sub></sup></th>
<th valign="bottom"><sup><sub>box<br/>AP</sub></sup></th>
<th valign="bottom"><sup><sub>mask<br/>AP</sub></sup></th>
<th valign="bottom"><sup><sub>model id</sub></sup></th>
<th valign="bottom"><sup><sub>download<br/>links</sub></sup></th>
<!-- TABLE BODY -->
<tr>
<td align="left"><sup><sub>R-50-FPN, GN</sub></sup></td>
<td align="left"><sup><sub>Mask R-CNN</sub></sup></td>
<td align="left"><sup><sub>2x</sub></sup></td>
<td align="right"><sup><sub>2</sub></sup></td>
<td align="right"><sup><sub>10.5</sub></sup></td>
<td align="right"><sup><sub>1.017</sub></sup></td>
<td align="right"><sup><sub>50.8</sub></sup></td>
<td align="right"><sup><sub>0.146 + 0.017</sub></sup></td>
<td align="right"><sup><sub>40.3</sub></sup></td>
<td align="right"><sup><sub>35.7</sub></sup></td>
<td align="right"><sup><sub>48616381</sub></sup></td>
<td align="left"><sup><sub>
<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl">model</a>
|
<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q/output/test/coco_2014_minival/generalized_rcnn/bbox_coco_2014_minival_results.json">boxes</a>
|
<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q/output/test/coco_2014_minival/generalized_rcnn/segmentations_coco_2014_minival_results.json">masks</a></sub></sup></td>
</tr>
<tr>
<td align="left"><sup><sub>R-101-FPN, GN</sub></sup></td>
<td align="left"><sup><sub>Mask R-CNN</sub></sup></td>
<td align="left"><sup><sub>2x</sub></sup></td>
<td align="right"><sup><sub>2</sub></sup></td>
<td align="right"><sup><sub>12.4</sub></sup></td>
<td align="right"><sup><sub>1.151</sub></sup></td>
<td align="right"><sup><sub>57.5</sub></sup></td>
<td align="right"><sup><sub>0.180 + 0.015</sub></sup></td>
<td align="right"><sup><sub>41.8</sub></sup></td>
<td align="right"><sup><sub>36.8</sub></sup></td>
<td align="right"><sup><sub>48616724</sub></sup></td>
<td align="left"><sup><sub>
<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616724/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_2x_gn_0416.13_26_34.GLnri4GR/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl">model</a>
|
<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616724/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_2x_gn_0416.13_26_34.GLnri4GR/output/test/coco_2014_minival/generalized_rcnn/bbox_coco_2014_minival_results.json">boxes</a>
|
<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616724/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_2x_gn_0416.13_26_34.GLnri4GR/output/test/coco_2014_minival/generalized_rcnn/segmentations_coco_2014_minival_results.json">masks</a></sub></sup></td>
</tr>
<!-- END E2E MASK RCNN GN TABLE -->
</tbody></table>
**Notes:**
- GN is applied on: (i) ResNet layers inherited from pre-training, (ii) the FPN-specific layers, (iii) the RoI bbox head, and (iv) the RoI mask head.
- These GN models use a 4conv+1fc RoI box head. The BN<sup>*</sup> counterpart with this head performs similarly with the default 2fc head: using this co
没有合适的资源?快使用搜索试试~ 我知道了~
温馨提示
【资源说明】 1、该资源包括项目的全部源码,下载可以直接使用! 2、本项目适合作为计算机、数学、电子信息等专业的课程设计、期末大作业和毕设项目,作为参考资料学习借鉴。 3、本资源作为“参考资料”如果需要实现其他功能,需要能看懂代码,并且热爱钻研,自行调试。 天池竞赛-津南数字制造算法挑战赛赛场二解决方案源码+项目说明.zip
资源推荐
资源详情
资源评论
收起资源包目录
天池竞赛-津南数字制造算法挑战赛赛场二解决方案源码+项目说明.zip (313个子文件)
zero_even_op.cc 1KB
Utils.cmake 11KB
Cuda.cmake 10KB
FindCuDNN.cmake 3KB
legacymake.cmake 2KB
Dependencies.cmake 2KB
Summary.cmake 2KB
Summary.cmake 2KB
zero_even_op.cu 2KB
Dockerfile 742B
.gitignore 560B
.gitignore 316B
.gitignore 71B
.gitignore 71B
.gitignore 71B
.gitkeep 0B
.gitkeep 0B
.gitkeep 0B
.gitkeep 0B
.gitkeep 0B
.gitkeep 0B
.gitkeep 0B
.gitkeep 0B
zlib-1.2.9.tar.gz 593KB
zero_even_op.h 1KB
33823288584_1d21cf0a26_k_example_output.jpg 997KB
17790319373_bd19b24cfc_k_example_output.jpg 961KB
learning_curve.jpg 520KB
gn.jpg 248KB
33823288584_1d21cf0a26_k.jpg 244KB
34501842524_3c858b3080_k.jpg 229KB
15673749081_767a7fa63a_k.jpg 220KB
18124840932_e42b3e377c_k.jpg 159KB
17790319373_bd19b24cfc_k.jpg 146KB
19064748793_bb942deea1_k.jpg 137KB
16004479832_a748d55f21_k.jpg 133KB
33887522274_eebd074106_k.jpg 119KB
24274813513_0cfd2ce6d0_k.jpg 111KB
LICENSE 10KB
voc_eval.m 1KB
xVOCap.m 258B
get_voc_opts.m 231B
Makefile 487B
MODEL_ZOO.md 116KB
README.md 17KB
INSTALL.md 9KB
README.md 8KB
README.md 7KB
GETTING_STARTED.md 6KB
ROUND1_README.md 6KB
FAQ.md 3KB
README.md 3KB
RECORD.md 2KB
CONTRIBUTING.md 1KB
README.md 729B
README.md 413B
CODE_OF_CONDUCT.md 244B
NOTICE 1KB
NOTICE 917B
config.py 46KB
test.py 34KB
model_builder.py 23KB
convert_pkl_to_pb.py 21KB
detector.py 21KB
data_augmentation_position_bak.py 21KB
FPN.py 20KB
json_dataset.py 19KB
json_dataset_evaluator.py 16KB
data_augmentation_position2.py 15KB
vis.py 15KB
test_engine.py 14KB
denseNet.py 14KB
task_evaluation.py 14KB
retinanet.py 13KB
boxes.py 13KB
net.py 12KB
model_convert_utils.py 12KB
retinanet_heads.py 12KB
loader.py 12KB
rpn.py 11KB
denseNet_transfer_learning.py 11KB
get_jinnan2_result_final.py 11KB
resNet50.py 11KB
fast_rcnn.py 11KB
get_jinnan2_result_final.py 11KB
ResNet.py 11KB
get_jinnan2_result_augmentation_mutil.py 11KB
resNet50_transfer_learning.py 11KB
get_jinnan2_result_augmentation.py 10KB
get_jinnan2_result_augmentation.py 10KB
data_augmentation_diffcult.py 10KB
mask_rcnn_heads.py 10KB
segms.py 10KB
rpn_generator.py 9KB
keypoints.py 9KB
generate_proposals.py 9KB
convert_cityscapes_to_coco.py 8KB
pickle_caffe_blobs.py 8KB
train.py 8KB
keypoint_rcnn_heads.py 8KB
共 313 条
- 1
- 2
- 3
- 4
资源评论
土豆片片
- 粉丝: 1520
- 资源: 5642
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功