Skip to content

Commit 8daec3b

Browse files
Tau-JTau-J
authored
Merge 0cdebc1 into 46a4169 
2 parents 46a4169 + 0cdebc1 commit 8daec3b

File tree

18 files changed

+1531
-67
lines changed

18 files changed

+1531
-67
lines changed

‎configs/_base_/datasets/lapa.py‎

Lines changed: 688 additions & 0 deletions
Large diffs are not rendered by default.

‎configs/face_2d_keypoint/rtmpose/README.md‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -22,3 +22,11 @@ Results on WFLW dataset
2222
| Model | Input Size | NME | Details and Download |
2323
| :-------: | :--------: | :--: | :---------------------------------------: |
2424
| RTMPose-m | 256x256 | 4.01 | [rtmpose_wflw.md](./wflw/rtmpose_wflw.md) |
25+
26+
### LaPa Dataset
27+
28+
Results on LaPa dataset
29+
30+
| Model | Input Size | NME | Details and Download |
31+
| :-------: | :--------: | :--: | :---------------------------------------: |
32+
| RTMPose-m | 256x256 | 1.29 | [rtmpose_lapa.md](./wflw/rtmpose_lapa.md) |

‎configs/face_2d_keypoint/rtmpose/lapa/rtmpose-m_8xb64-120e_lapa-256x256.py‎

Lines changed: 247 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,247 @@
1+
_base_ = ['../../../_base_/default_runtime.py']
2+
3+
# runtime
4+
max_epochs = 120
5+
stage2_num_epochs = 10
6+
base_lr = 4e-3
7+
8+
train_cfg = dict(max_epochs=max_epochs, val_interval=1)
9+
randomness = dict(seed=21)
10+
11+
# optimizer
12+
optim_wrapper = dict(
13+
type='OptimWrapper',
14+
optimizer=dict(type='AdamW', lr=base_lr, weight_decay=0.05),
15+
paramwise_cfg=dict(
16+
norm_decay_mult=0, bias_decay_mult=0, bypass_duplicate=True))
17+
18+
# learning rate
19+
param_scheduler = [
20+
dict(
21+
type='LinearLR',
22+
start_factor=1.0e-5,
23+
by_epoch=False,
24+
begin=0,
25+
end=1000),
26+
dict(
27+
# use cosine lr from 150 to 300 epoch
28+
type='CosineAnnealingLR',
29+
eta_min=base_lr * 0.05,
30+
begin=max_epochs // 2,
31+
end=max_epochs,
32+
T_max=max_epochs // 2,
33+
by_epoch=True,
34+
convert_to_iter_based=True),
35+
]
36+
37+
# automatically scaling LR based on the actual training batch size
38+
auto_scale_lr = dict(base_batch_size=512)
39+
40+
# codec settings
41+
codec = dict(
42+
type='SimCCLabel',
43+
input_size=(256, 256),
44+
sigma=(5.66, 5.66),
45+
simcc_split_ratio=2.0,
46+
normalize=False,
47+
use_dark=False)
48+
49+
# model settings
50+
model = dict(
51+
type='TopdownPoseEstimator',
52+
data_preprocessor=dict(
53+
type='PoseDataPreprocessor',
54+
mean=[123.675, 116.28, 103.53],
55+
std=[58.395, 57.12, 57.375],
56+
bgr_to_rgb=True),
57+
backbone=dict(
58+
_scope_='mmdet',
59+
type='CSPNeXt',
60+
arch='P5',
61+
expand_ratio=0.5,
62+
deepen_factor=0.67,
63+
widen_factor=0.75,
64+
out_indices=(4, ),
65+
channel_attention=True,
66+
norm_cfg=dict(type='SyncBN'),
67+
act_cfg=dict(type='SiLU'),
68+
init_cfg=dict(
69+
type='Pretrained',
70+
prefix='backbone.',
71+
checkpoint='https://download.openmmlab.com/mmpose/v1/projects/'
72+
'rtmposev1/cspnext-m_udp-aic-coco_210e-256x192-f2f7d6f6_20230130.pth' # noqa
73+
)),
74+
head=dict(
75+
type='RTMCCHead',
76+
in_channels=768,
77+
out_channels=106,
78+
input_size=codec['input_size'],
79+
in_featuremap_size=(8, 8),
80+
simcc_split_ratio=codec['simcc_split_ratio'],
81+
final_layer_kernel_size=7,
82+
gau_cfg=dict(
83+
hidden_dims=256,
84+
s=128,
85+
expansion_factor=2,
86+
dropout_rate=0.,
87+
drop_path=0.,
88+
act_fn='SiLU',
89+
use_rel_bias=False,
90+
pos_enc=False),
91+
loss=dict(
92+
type='KLDiscretLoss',
93+
use_target_weight=True,
94+
beta=10.,
95+
label_softmax=True),
96+
decoder=codec),
97+
test_cfg=dict(flip_test=True, ))
98+
99+
# base dataset settings
100+
dataset_type = 'LapaDataset'
101+
data_mode = 'topdown'
102+
data_root = 'data/LaPa/'
103+
104+
backend_args = dict(backend='local')
105+
# backend_args = dict(
106+
# backend='petrel',
107+
# path_mapping=dict({
108+
# f'{data_root}': 's3://openmmlab/datasets/pose/LaPa/',
109+
# f'{data_root}': 's3://openmmlab/datasets/pose/LaPa/'
110+
# }))
111+
112+
# pipelines
113+
train_pipeline = [
114+
dict(type='LoadImage', backend_args=backend_args),
115+
dict(type='GetBBoxCenterScale'),
116+
dict(type='RandomFlip', direction='horizontal'),
117+
dict(type='RandomHalfBody'),
118+
dict(
119+
type='RandomBBoxTransform', scale_factor=[0.5, 1.5], rotate_factor=80),
120+
dict(type='TopdownAffine', input_size=codec['input_size']),
121+
dict(type='mmdet.YOLOXHSVRandomAug'),
122+
dict(type='PhotometricDistortion'),
123+
dict(
124+
type='Albumentation',
125+
transforms=[
126+
dict(type='Blur', p=0.2),
127+
dict(type='MedianBlur', p=0.2),
128+
dict(
129+
type='CoarseDropout',
130+
max_holes=1,
131+
max_height=0.4,
132+
max_width=0.4,
133+
min_holes=1,
134+
min_height=0.2,
135+
min_width=0.2,
136+
p=1.0),
137+
]),
138+
dict(type='GenerateTarget', encoder=codec),
139+
dict(type='PackPoseInputs')
140+
]
141+
val_pipeline = [
142+
dict(type='LoadImage', backend_args=backend_args),
143+
dict(type='GetBBoxCenterScale'),
144+
dict(type='TopdownAffine', input_size=codec['input_size']),
145+
dict(type='PackPoseInputs')
146+
]
147+
148+
train_pipeline_stage2 = [
149+
dict(type='LoadImage', backend_args=backend_args),
150+
dict(type='GetBBoxCenterScale'),
151+
dict(type='RandomFlip', direction='horizontal'),
152+
# dict(type='RandomHalfBody'),
153+
dict(
154+
type='RandomBBoxTransform',
155+
shift_factor=0.,
156+
scale_factor=[0.75, 1.25],
157+
rotate_factor=60),
158+
dict(type='TopdownAffine', input_size=codec['input_size']),
159+
dict(type='mmdet.YOLOXHSVRandomAug'),
160+
dict(
161+
type='Albumentation',
162+
transforms=[
163+
dict(type='Blur', p=0.1),
164+
dict(type='MedianBlur', p=0.1),
165+
dict(
166+
type='CoarseDropout',
167+
max_holes=1,
168+
max_height=0.4,
169+
max_width=0.4,
170+
min_holes=1,
171+
min_height=0.2,
172+
min_width=0.2,
173+
p=0.5),
174+
]),
175+
dict(type='GenerateTarget', encoder=codec),
176+
dict(type='PackPoseInputs')
177+
]
178+
179+
# data loaders
180+
train_dataloader = dict(
181+
batch_size=32,
182+
num_workers=10,
183+
persistent_workers=True,
184+
sampler=dict(type='DefaultSampler', shuffle=True),
185+
dataset=dict(
186+
type=dataset_type,
187+
data_root=data_root,
188+
data_mode=data_mode,
189+
ann_file='annotations/lapa_train.json',
190+
data_prefix=dict(img='train/images/'),
191+
pipeline=train_pipeline,
192+
))
193+
val_dataloader = dict(
194+
batch_size=32,
195+
num_workers=10,
196+
persistent_workers=True,
197+
drop_last=False,
198+
sampler=dict(type='DefaultSampler', shuffle=False, round_up=False),
199+
dataset=dict(
200+
type=dataset_type,
201+
data_root=data_root,
202+
data_mode=data_mode,
203+
ann_file='annotations/lapa_val.json',
204+
data_prefix=dict(img='val/images/'),
205+
test_mode=True,
206+
pipeline=val_pipeline,
207+
))
208+
test_dataloader = dict(
209+
batch_size=32,
210+
num_workers=10,
211+
persistent_workers=True,
212+
drop_last=False,
213+
sampler=dict(type='DefaultSampler', shuffle=False, round_up=False),
214+
dataset=dict(
215+
type=dataset_type,
216+
data_root=data_root,
217+
data_mode=data_mode,
218+
ann_file='annotations/lapa_test.json',
219+
data_prefix=dict(img='test/images/'),
220+
test_mode=True,
221+
pipeline=val_pipeline,
222+
))
223+
224+
# hooks
225+
default_hooks = dict(
226+
checkpoint=dict(
227+
save_best='NME', rule='less', max_keep_ckpts=1, interval=1))
228+
229+
custom_hooks = [
230+
dict(
231+
type='EMAHook',
232+
ema_type='ExpMomentumEMA',
233+
momentum=0.0002,
234+
update_buffers=True,
235+
priority=49),
236+
dict(
237+
type='mmdet.PipelineSwitchHook',
238+
switch_epoch=max_epochs - stage2_num_epochs,
239+
switch_pipeline=train_pipeline_stage2)
240+
]
241+
242+
# evaluators
243+
val_evaluator = dict(
244+
type='NME',
245+
norm_mode='keypoint_distance',
246+
)
247+
test_evaluator = val_evaluator

‎configs/face_2d_keypoint/rtmpose/lapa/rtmpose_lapa.md‎

Lines changed: 40 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,40 @@
1+
<!-- [ALGORITHM] -->
2+
3+
<details>
4+
<summary align="right"><a href="https://arxiv.org/abs/2212.07784">RTMDet (ArXiv 2022)</a></summary>
5+
6+
```bibtex
7+
@misc{lyu2022rtmdet,
8+
title={RTMDet: An Empirical Study of Designing Real-Time Object Detectors},
9+
author={Chengqi Lyu and Wenwei Zhang and Haian Huang and Yue Zhou and Yudong Wang and Yanyi Liu and Shilong Zhang and Kai Chen},
10+
year={2022},
11+
eprint={2212.07784},
12+
archivePrefix={arXiv},
13+
primaryClass={cs.CV}
14+
}
15+
```
16+
17+
</details>
18+
19+
<!-- [DATASET] -->
20+
21+
<details>
22+
<summary align="right"><a href="https://aaai.org/ojs/index.php/AAAI/article/view/6832/6686">LaPa (AAAI'2020)</a></summary>
23+
24+
```bibtex
25+
@inproceedings{liu2020new,
26+
title={A New Dataset and Boundary-Attention Semantic Segmentation for Face Parsing.},
27+
author={Liu, Yinglu and Shi, Hailin and Shen, Hao and Si, Yue and Wang, Xiaobo and Mei, Tao},
28+
booktitle={AAAI},
29+
pages={11637--11644},
30+
year={2020}
31+
}
32+
```
33+
34+
</details>
35+
36+
Results on COCO-WholeBody-Face val set
37+
38+
| Arch | Input Size | NME | ckpt | log |
39+
| :------------------------------------------------------------- | :--------: | :--: | :------------------------------------------------------------: | :------------------------------------------------------------: |
40+
| [pose_rtmpose_m](/configs/face_2d_keypoint/rtmpose/lapa/rtmpose-m_8xb64-120e_lapa-256x256.py) | 256x256 | 1.29 | [ckpt](https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/rtmpose-m_simcc-lapa_pt-aic-coco_120e-256x256-762b1ae2_20230422.pth) | [log](https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/rtmpose-m_simcc-lapa_pt-aic-coco_120e-256x256-762b1ae2_20230422.json) |

‎configs/face_2d_keypoint/rtmpose/lapa/rtmpose_lapa.yml‎

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,15 @@
1+
Models:
2+
- Config: configs/face_2d_keypoint/rtmpose/lapa/rtmpose-m_8xb64-120e_lapa-256x256.py
3+
In Collection: RTMPose
4+
Alias: face
5+
Metadata:
6+
Architecture:
7+
- RTMPose
8+
Training Data: LaPa
9+
Name: rtmpose-m_8xb64-120e_lapa-256x256
10+
Results:
11+
- Dataset: WFLW
12+
Metrics:
13+
NME: 1.29
14+
Task: Face 2D Keypoint
15+
Weights: https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/rtmpose-m_simcc-lapa_pt-aic-coco_120e-256x256-762b1ae2_20230422.pth

‎docs/en/dataset_zoo/2d_face_keypoint.md‎

Lines changed: 56 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,7 @@ MMPose supported datasets:
1010
- [AFLW](#aflw-dataset) \[ [Homepage](https://www.tugraz.at/institute/icg/research/team-bischof/lrs/downloads/aflw/) \]
1111
- [COFW](#cofw-dataset) \[ [Homepage](http://www.vision.caltech.edu/xpburgos/ICCV13/) \]
1212
- [COCO-WholeBody-Face](#coco-wholebody-face) \[ [Homepage](https://github.com/jin-s13/COCO-WholeBody/) \]
13+
- [LaPa](#lapa-dataset) \[ [Homepage](https://github.com/JDAI-CV/lapa-dataset) \]
1314

1415
## 300W Dataset
1516

@@ -325,3 +326,58 @@ mmpose
325326
Please also install the latest version of [Extended COCO API](https://github.com/jin-s13/xtcocoapi) to support COCO-WholeBody evaluation:
326327

327328
`pip install xtcocotools`
329+
330+
## LaPa
331+
332+
<!-- [DATASET] -->
333+
334+
<details>
335+
<summary align="right"><a href="https://aaai.org/ojs/index.php/AAAI/article/view/6832/6686">LaPa (AAAI'2020)</a></summary>
336+
337+
```bibtex
338+
@inproceedings{liu2020new,
339+
title={A New Dataset and Boundary-Attention Semantic Segmentation for Face Parsing.},
340+
author={Liu, Yinglu and Shi, Hailin and Shen, Hao and Si, Yue and Wang, Xiaobo and Mei, Tao},
341+
booktitle={AAAI},
342+
pages={11637--11644},
343+
year={2020}
344+
}
345+
```
346+
347+
</details>
348+
349+
<div align="center">
350+
<img src="https://github.com/lucia123/lapa-dataset/raw/master/sample.png" height="200px">
351+
</div>
352+
353+
For [LaPa](https://github.com/JDAI-CV/lapa-dataset) dataset, images can be downloaded from [their github page](https://github.com/JDAI-CV/lapa-dataset).
354+
355+
Download and extract them under $MMPOSE/data, and use our `tools/dataset_converters/lapa2coco.py` to make them look like this:
356+
357+
```text
358+
mmpose
359+
├── mmpose
360+
├── docs
361+
├── tests
362+
├── tools
363+
├── configs
364+
`── data
365+
│── LaPa
366+
│-- annotations
367+
│ │-- lapa_train.json
368+
│ |-- lapa_val.json
369+
│ |-- lapa_test.json
370+
│-- train
371+
│ │-- images
372+
│ │-- labels
373+
│ │-- landmarks
374+
│-- val
375+
│ │-- images
376+
│ │-- labels
377+
│ │-- landmarks
378+
`-- test
379+
│ │-- images
380+
│ │-- labels
381+
│ │-- landmarks
382+
383+
```

‎mmpose/datasets/datasets/face/__init__.py‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -3,9 +3,10 @@
33
from .coco_wholebody_face_dataset import CocoWholeBodyFaceDataset
44
from .cofw_dataset import COFWDataset
55
from .face_300w_dataset import Face300WDataset
6+
from .lapa_dataset import LapaDataset
67
from .wflw_dataset import WFLWDataset
78

89
__all__ = [
910
'Face300WDataset', 'WFLWDataset', 'AFLWDataset', 'COFWDataset',
10-
'CocoWholeBodyFaceDataset'
11+
'CocoWholeBodyFaceDataset', 'LapaDataset'
1112
]

0 commit comments

Comments
 (0)