Skip to content

Commit 4e40f6f

Browse files
joe-siyuan-qiaoxvjiarui
joe-siyuan-qiao
and
xvjiarui
authored
Add support for DetectoRS. (open-mmlab#3064)
* Add support for DetectoRS. * Fix Travis CI build failure and remove loading rfp_pretrained for test * Revise the code according to the PR reviews: * remove unnecessary `,` * add docstring to ASPP and ConvAWS2d * make args of ASPP configurable * simplify the code of ConvAWS2d * import nn and functional for SAConv2d * rewrite rfp_conv with build_conv_layer and constant_init * rename rfp_inp with rfp_inplanes * reformat code to pass Travis CI build * Revise codes according to the PR reviews: * Create new backbone DetectoRS_ResNe(X)t to incorporate RFP and SAC. * Initialize modules with mmcv.cnn. * Revise the codes of RFP. * Revise codes according to the PR comments: * Add docstrings to Bottleneck in mmdet/models/backbones/detectors_resnet.py. * Add docstrings to _load_from_state_dict in mmdet/ops/conv_ws.py. * Initialize the switch of SAC with mmcv. * minior refactor and docstring * update doc * update README * move init into init_weights() * update model zoo * add gfl Co-authored-by: Jiarui XU <[email protected]>
1 parent 39e6257 commit 4e40f6f

18 files changed

+962
-3
lines changed

‎README.md

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -92,6 +92,8 @@ Supported methods:
9292
- [x] [Mixed Precision (FP16) Training](configs/fp16/README.md)
9393
- [x] [InstaBoost](configs/instaboost/README.md)
9494
- [x] [GRoIE](configs/groie/README.md)
95+
- [x] [DetectoRS](configs/detectors/README.md)
96+
- [x] [Generalized Focal Loss](configs/gfl/README.md)
9597

9698
Some other methods are also supported in [projects using MMDetection](./docs/projects.md).
9799

‎configs/detectors/README.md

Lines changed: 37 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,37 @@
1+
# DetectoRS
2+
3+
## Introduction
4+
5+
We provide the config files for [DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution](https://arxiv.org/pdf/2006.02334.pdf).
6+
7+
```BibTeX
8+
@article{qiao2020detectors,
9+
title={DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution},
10+
author={Qiao, Siyuan and Chen, Liang-Chieh and Yuille, Alan},
11+
journal={arXiv preprint arXiv:2006.02334},
12+
year={2020}
13+
}
14+
```
15+
16+
## Results and Models
17+
18+
DetectoRS includes two major components:
19+
20+
- Recursive Feature Pyramid (RFP).
21+
- Switchable Atrous Convolution (SAC).
22+
23+
They can be used independently.
24+
Combining them together results in DetectoRS.
25+
The results on COCO 2017 val are shown in the below table.
26+
27+
| Method | Detector | Lr schd | Mem (GB) | Inf time (fps) | box AP | mask AP | Download |
28+
|:------:|:--------:|:-------:|:--------:|:--------------:|:------:|:-------:|:--------:|
29+
| RFP | Cascade + ResNet-50 | 1x | 7.5 | - | 44.8 | | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/cascade_rcnn_r50_rfp_1x_coco/cascade_rcnn_r50_rfp_1x_coco-8cf51bfd.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/cascade_rcnn_r50_rfp_1x_coco/cascade_rcnn_r50_rfp_1x_coco_20200624_104126.log.json) |
30+
| SAC | Cascade + ResNet-50 | 1x | 5.6 | - | 45.0| | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/cascade_rcnn_r50_sac_1x_coco/cascade_rcnn_r50_sac_1x_coco-24bfda62.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/cascade_rcnn_r50_sac_1x_coco/cascade_rcnn_r50_sac_1x_coco_20200624_104402.log.json) |
31+
| DetectoRS | Cascade + ResNet-50 | 1x | 9.9 | - | 46.9 | | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/detectors_cascade_rcnn_r50_1x_coco/detectors_cascade_rcnn_r50_1x_coco-0db1ab6a.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/detectors_cascade_rcnn_r50_1x_coco/detectors_cascade_rcnn_r50_1x_coco_20200624_103448.log.json) |
32+
| RFP | HTC + ResNet-50 | 1x | 11.2 | - | 46.6 | 40.9 | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/htc_r50_rfp_1x_coco/htc_r50_rfp_1x_coco-8ff87c51.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/htc_r50_rfp_1x_coco/htc_r50_rfp_1x_coco_20200624_103053.log.json) |
33+
| SAC | HTC + ResNet-50 | 1x | 9.3 | - | 46.4 | 40.9 | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/htc_r50_sac_1x_coco/htc_r50_sac_1x_coco-bfa60c54.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/htc_r50_sac_1x_coco/htc_r50_sac_1x_coco_20200624_103111.log.json) |
34+
| DetectoRS | HTC + ResNet-50 | 1x | 13.6 | - | 49.1 | 42.6 | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/detectors_htc_r50_1x_coco/detectors_htc_r50_1x_coco-329b1453.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/detectors/detectors_htc_r50_1x_coco/detectors_htc_r50_1x_coco_20200624_103659.log.json) |
35+
36+
*Note*: This is a re-implementation based on MMDetection-V2.
37+
The original implementation is based on MMDetection-V1.

‎configs/detectors/cascade_rcnn_r50_rfp_1x_coco.py

Lines changed: 28 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,28 @@
1+
_base_ = [
2+
'../_base_/models/cascade_rcnn_r50_fpn.py',
3+
'../_base_/datasets/coco_detection.py',
4+
'../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
5+
]
6+
7+
model = dict(
8+
backbone=dict(
9+
type='DetectoRS_ResNet',
10+
conv_cfg=dict(type='ConvAWS'),
11+
output_img=True),
12+
neck=dict(
13+
type='RFP',
14+
rfp_steps=2,
15+
aspp_out_channels=64,
16+
aspp_dilations=(1, 3, 6, 1),
17+
rfp_backbone=dict(
18+
rfp_inplanes=256,
19+
type='DetectoRS_ResNet',
20+
depth=50,
21+
num_stages=4,
22+
out_indices=(0, 1, 2, 3),
23+
frozen_stages=1,
24+
norm_cfg=dict(type='BN', requires_grad=True),
25+
norm_eval=True,
26+
conv_cfg=dict(type='ConvAWS'),
27+
pretrained='torchvision://resnet50',
28+
style='pytorch')))

‎configs/detectors/cascade_rcnn_r50_sac_1x_coco.py

Lines changed: 12 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,12 @@
1+
_base_ = [
2+
'../_base_/models/cascade_rcnn_r50_fpn.py',
3+
'../_base_/datasets/coco_detection.py',
4+
'../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
5+
]
6+
7+
model = dict(
8+
backbone=dict(
9+
type='DetectoRS_ResNet',
10+
conv_cfg=dict(type='ConvAWS'),
11+
sac=dict(type='SAC', use_deform=True),
12+
stage_with_sac=(False, True, True, True)))

‎configs/detectors/detectors_cascade_rcnn_r50_1x_coco.py

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,32 @@
1+
_base_ = [
2+
'../_base_/models/cascade_rcnn_r50_fpn.py',
3+
'../_base_/datasets/coco_detection.py',
4+
'../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
5+
]
6+
7+
model = dict(
8+
backbone=dict(
9+
type='DetectoRS_ResNet',
10+
conv_cfg=dict(type='ConvAWS'),
11+
sac=dict(type='SAC', use_deform=True),
12+
stage_with_sac=(False, True, True, True),
13+
output_img=True),
14+
neck=dict(
15+
type='RFP',
16+
rfp_steps=2,
17+
aspp_out_channels=64,
18+
aspp_dilations=(1, 3, 6, 1),
19+
rfp_backbone=dict(
20+
rfp_inplanes=256,
21+
type='DetectoRS_ResNet',
22+
depth=50,
23+
num_stages=4,
24+
out_indices=(0, 1, 2, 3),
25+
frozen_stages=1,
26+
norm_cfg=dict(type='BN', requires_grad=True),
27+
norm_eval=True,
28+
conv_cfg=dict(type='ConvAWS'),
29+
sac=dict(type='SAC', use_deform=True),
30+
stage_with_sac=(False, True, True, True),
31+
pretrained='torchvision://resnet50',
32+
style='pytorch')))

‎configs/detectors/detectors_htc_r50_1x_coco.py

Lines changed: 28 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,28 @@
1+
_base_ = '../htc/htc_r50_fpn_1x_coco.py'
2+
3+
model = dict(
4+
backbone=dict(
5+
type='DetectoRS_ResNet',
6+
conv_cfg=dict(type='ConvAWS'),
7+
sac=dict(type='SAC', use_deform=True),
8+
stage_with_sac=(False, True, True, True),
9+
output_img=True),
10+
neck=dict(
11+
type='RFP',
12+
rfp_steps=2,
13+
aspp_out_channels=64,
14+
aspp_dilations=(1, 3, 6, 1),
15+
rfp_backbone=dict(
16+
rfp_inplanes=256,
17+
type='DetectoRS_ResNet',
18+
depth=50,
19+
num_stages=4,
20+
out_indices=(0, 1, 2, 3),
21+
frozen_stages=1,
22+
norm_cfg=dict(type='BN', requires_grad=True),
23+
norm_eval=True,
24+
conv_cfg=dict(type='ConvAWS'),
25+
sac=dict(type='SAC', use_deform=True),
26+
stage_with_sac=(False, True, True, True),
27+
pretrained='torchvision://resnet50',
28+
style='pytorch')))

‎configs/detectors/htc_r50_rfp_1x_coco.py

Lines changed: 24 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,24 @@
1+
_base_ = '../htc/htc_r50_fpn_1x_coco.py'
2+
3+
model = dict(
4+
backbone=dict(
5+
type='DetectoRS_ResNet',
6+
conv_cfg=dict(type='ConvAWS'),
7+
output_img=True),
8+
neck=dict(
9+
type='RFP',
10+
rfp_steps=2,
11+
aspp_out_channels=64,
12+
aspp_dilations=(1, 3, 6, 1),
13+
rfp_backbone=dict(
14+
rfp_inplanes=256,
15+
type='DetectoRS_ResNet',
16+
depth=50,
17+
num_stages=4,
18+
out_indices=(0, 1, 2, 3),
19+
frozen_stages=1,
20+
norm_cfg=dict(type='BN', requires_grad=True),
21+
norm_eval=True,
22+
conv_cfg=dict(type='ConvAWS'),
23+
pretrained='torchvision://resnet50',
24+
style='pytorch')))

‎configs/detectors/htc_r50_sac_1x_coco.py

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,8 @@
1+
_base_ = '../htc/htc_r50_fpn_1x_coco.py'
2+
3+
model = dict(
4+
backbone=dict(
5+
type='DetectoRS_ResNet',
6+
conv_cfg=dict(type='ConvAWS'),
7+
sac=dict(type='SAC', use_deform=True),
8+
stage_with_sac=(False, True, True, True)))

‎docs/model_zoo.md

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -138,6 +138,12 @@ Please refer to [Dynamic R-CNN](https://github.com/open-mmlab/mmdetection/blob/m
138138
### PointRend
139139
Please refer to [PointRend](https://github.com/open-mmlab/mmdetection/blob/master/configs/point_rend) for details.
140140

141+
### DetectoRS
142+
Please refer to [DetectoRS](https://github.com/open-mmlab/mmdetection/blob/master/configs/detectors) for details.
143+
144+
### Generalized Focal Loss
145+
Please refer to [Generalized Focal Loss](https://github.com/open-mmlab/mmdetection/blob/master/configs/gfl) for details.
146+
141147
### Other datasets
142148

143149
We also benchmark some methods on [PASCAL VOC](https://github.com/open-mmlab/mmdetection/blob/master/configs/pascal_voc), [Cityscapes](https://github.com/open-mmlab/mmdetection/blob/master/configs/cityscapes) and [WIDER FACE](https://github.com/open-mmlab/mmdetection/blob/master/configs/wider_face).

‎mmdet/models/backbones/__init__.py

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,5 @@
1+
from .detectors_resnet import DetectoRS_ResNet
2+
from .detectors_resnext import DetectoRS_ResNeXt
13
from .hourglass import HourglassNet
24
from .hrnet import HRNet
35
from .regnet import RegNet
@@ -8,5 +10,5 @@
810

911
__all__ = [
1012
'RegNet', 'ResNet', 'ResNetV1d', 'ResNeXt', 'SSDVGG', 'HRNet', 'Res2Net',
11-
'HourglassNet'
13+
'HourglassNet', 'DetectoRS_ResNet', 'DetectoRS_ResNeXt'
1214
]

0 commit comments

Comments
 (0)