milk333445/PoseEstimationTutorial

MMpose、MMdetecetion安裝流程、RTMpose訓練流程

mmpose

用mim安裝mmcv

pip install -U openmim
mim install mmengine
mim install mmcv

下載mmpose

git clone https://github.com/open-mmlab/mmpose.git
mim install "mmpose>=1.1.0"

檢查有沒有安裝成功

# 检查 mmcv
import mmcv
from mmcv.ops import get_compiling_cuda_version, get_compiler_version
print('MMCV版本', mmcv.__version__)
print('CUDA版本', get_compiling_cuda_version())
print('編譯器版本', get_compiler_version())

# 檢查 mmpose
import mmpose
print('mmpose版本', mmpose.__version__)

mmdetection

下載mmdetection

git clone https://github.com/open-mmlab/mmdetection.git
mim install "mmdet>=3.1.0"

檢查有沒有安裝成功

# 檢查 mmpose
import mmdet
print('mmdetection版本', mmdet.__version__)

使用命令行進行預測

單張圖像進行預測

faster_rcnn + hrnet

python demo/topdown_demo_with_mmdet.py \ 
        demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py \# 目標檢測模型的config文件
        https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \ # 目標檢測模型的模型權重文件
        configs/body_2d_keypoint/topdown_heatmap/coco/td-hm_hrnet-w32_8xb64-210e_coco-256x192.py \ # 關鍵點檢測的模型config文件
        https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth \# 關鍵點檢測模型的模型權重文件
        --input data/test/multi-person.jpeg \ # 要檢測的圖像或影片路徑
        --output-root outputs/B1_HRNet_1 \ # 模型預測結果保存位置
        --device cuda:0 \ # 指定gpu，也可以用cpu
        --bbox-thr 0.5 \ # 框檢測的置信度thredhold
        --kpt-thr 0.2 \ # 關鍵點檢測的置信度threshold
        --nms-thr 0.3 \# 框的nms iou threshold
        --radius 8 \ # 可視化關鍵點半徑
        --thickness 4 \ # 框的線寬
        --draw-bbox \ # 是否要畫框
        --draw-heatmap \ # 是否要化關鍵點熱力圖
        --show-kpt-idx  # 是否在點上標明是第幾號關鍵點

faster_rcnn + rtmpose

python demo/topdown_demo_with_mmdet.py \
        demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py \
        https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
        projects/rtmpose/rtmpose/body_2d_keypoint/rtmpose-s_8xb256-420e_coco-256x192.py \
        https://download.openmmlab.com/mmpose/v1/projects/rtmpose/rtmpose-s_simcc-aic-coco_pt-aic-coco_420e-256x192-fcb2599b_20230126.pth \
        --input data/test/multi-person.jpeg \
        --output-root outputs/B1_RTM_1 \
        --device cuda:0 \
        --bbox-thr 0.5 \
        --kpt-thr 0.5 \
        --nms-thr 0.3 \
        --radius 8 \
        --thickness 4 \
        --draw-bbox \
        --draw-heatmap \
        --show-kpt-idx

對影片處理

faster_rcnn + hrnet

python demo/topdown_demo_with_mmdet.py \
        demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py \
        https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
        configs/body_2d_keypoint/topdown_heatmap/coco/td-hm_hrnet-w32_8xb64-210e_coco-256x192.py \
        https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth \
        --input data/test/mother_wx.mp4 \
        --output-root outputs/B1_HRNet_2 \
        --device cuda:0 \
        --bbox-thr 0.5 \
        --kpt-thr 0.2 \
        --nms-thr 0.3 \
        --radius 5 \
        --thickness 2 \
        --draw-bbox \
        --draw-heatmap \
        --show-kpt-idx

faster_rcnn + rtmpose

python demo/topdown_demo_with_mmdet.py \
        demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py \
        https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
        projects/rtmpose/rtmpose/body_2d_keypoint/rtmpose-s_8xb256-420e_coco-256x192.py \
        https://download.openmmlab.com/mmpose/v1/projects/rtmpose/rtmpose-s_simcc-aic-coco_pt-aic-coco_420e-256x192-fcb2599b_20230126.pth \
        --input data/test/mother_wx.mp4 \
        --output-root outputs/B1_RTM_2 \
        --device cuda:0 \
        --bbox-thr 0.5 \
        --kpt-thr 0.5 \
        --nms-thr 0.3 \
        --radius 5 \
        --thickness 2 \
        --draw-bbox \
        --draw-heatmap \
        --show-kpt-idx

用python api調用

• 要記得執行的py檔要跟mmdetection、mmpose平行，不能放在裡面，會有找不到包的問題

導入必要的函數和類

import cv2
import os
from PIL import Image
import matplotlib.pyplot as plt
import torch
import mmcv
import numpy as np
from mmcv import imread
import mmengine
from mmengine.registry import init_default_scope

from mmpose.apis import inference_topdown # 姿態檢測推論
from mmpose.apis import init_model as init_pose_estimator # 初始化姿態檢測模型
from mmpose.evaluation.functional import nms
from mmpose.registry import VISUALIZERS
from mmpose.structures import merge_data_samples

from mmdet.apis import inference_detector # 推論目標檢測模型需要
from mmdet.apis import init_detector # 初始化目標檢測模型
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# 檢測圖像
img_path = './mmpose/data/test/multi-person.jpeg'

建構目標檢測模型

# 建構目標檢測模型
## faster-rcnn
detector = init_detector(
    'mmpose/demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py', # 目標檢測模型config
    'https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth', # 對應權重
    device=device
)

建構姿態檢測模型

pose_estimator = init_pose_estimator(
    'mmpose/configs/body_2d_keypoint/topdown_heatmap/coco/td-hm_hrnet-w32_8xb64-210e_coco-256x192.py',
    'https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth',
    device=device,
    cfg_options={'model': {'test_cfg': {'output_heatmaps': True}}}
)

開始推論目標檢測模型

init_default_scope(detector.cfg.get('default_scope', 'mmdet'))
# 預測框
detect_result = inference_detector(detector, img_path)

• detect_result裡面有三種結果
  • detect_result.pred_instances.labels:預測出來的框的類別
  • detect_result.pred_instances.scores:預測出來的置信度
  • detect_result.pred_instances.bboxes:預測出來的框座標

# 預測出來的框的類別
tensor([ 0,  0,  0,  0,  0,  0,  0,  0,  0, 11,  0, 58,  0,  0, 58,  0,  0,  2,
         0,  0,  2,  0,  0,  0,  0, 24,  2, 58, 26, 26,  0,  0, 58,  0, 26,  0,
        27, 24,  0, 26, 26, 26, 26, 26, 58,  0, 26, 26, 26,  2, 32,  2, 58, 39,
         2,  0, 58, 24, 58, 32, 24, 26,  0,  0,  2, 37,  0, 39,  9, 74,  0,  2,
        58, 24,  0,  0, 26, 29], device='cuda:0')

# 預測出來的置信度
tensor([0.9994, 0.9988, 0.9974, 0.9963, 0.9959, 0.9930, 0.9914, 0.9901, 0.9868,
        0.9824, 0.9632, 0.9272, 0.9080, 0.8950, 0.8794, 0.8582, 0.8248, 0.8189,
        0.7240, 0.7185, 0.7055, 0.6654, 0.6203, 0.6010, 0.6003, 0.5710, 0.5653,
        0.5058, 0.4549, 0.4088, 0.3942, 0.3909, 0.3533, 0.3103, 0.3030, 0.2594,
        0.2292, 0.2223, 0.2156, 0.1888, 0.1866, 0.1847, 0.1792, 0.1770, 0.1716,
        0.1673, 0.1398, 0.1391, 0.1390, 0.1373, 0.1348, 0.1251, 0.1205, 0.1156,
        0.1132, 0.1121, 0.1010, 0.1009, 0.0940, 0.0921, 0.0919, 0.0914, 0.0885,
        0.0821, 0.0805, 0.0724, 0.0697, 0.0687, 0.0675, 0.0630, 0.0626, 0.0624,
        0.0587, 0.0578, 0.0557, 0.0536, 0.0510, 0.0502], device='cuda:0')

# 預測出來的框座標
tensor([[1705.6440,   96.2784, 2114.5181, 1301.1793],
        [1267.5966,   94.7074, 1701.0554, 1316.2964],
        [ 720.3899,  172.2511, 1152.6925, 1267.8408],
        [   7.2623,  238.2423,  170.9962, 1140.8058],
        [1063.1945,  219.2218, 1348.9850, 1244.7227],
        [ 436.2329,  428.5780,  586.0933,  911.8918],
        [ 110.0086,  212.1305,  324.6433, 1123.0087],
        [2079.3340,  478.2407, 2224.4050,  925.2119],
        [ 543.7142,  406.4844,  650.8373,  919.4562],
        [ 514.5025,  341.8667,  580.5118,  430.7810],

先對框的置信度篩選以及過nms

CONF_THRES = 0.5
pred_instance = detect_result.pred_instances.cpu().numpy()
bboxes = np.concatenate(
    (pred_instance.bboxes, pred_instance.scores[:, None]),
    axis=1
)
bboxes = bboxes[np.logical_and(pred_instance.labels == 0, pred_instance.scores > CONF_THRES)]
bboxes = bboxes[nms(bboxes, 0.3)][:, :4]

[[1705.644       96.27843   2114.518     1301.1793   ]
 [1267.5966      94.70741   1701.0554    1316.2964   ]
 [ 720.3899     172.25111   1152.6925    1267.8408   ]
 [   7.2623434  238.24232    170.99622   1140.8058   ]
 [1063.1945     219.22185   1348.985     1244.7227   ]
 [ 436.23294    428.578      586.0933     911.8918   ]
 [ 110.008575   212.1305     324.64334   1123.0087   ]
 [2079.334      478.2407    2224.405      925.2119   ]
 [ 543.71423    406.48438    650.8373     919.45624  ]
 [2363.2996     501.87018   2455.2249     857.3131   ]
 [2308.198      561.452     2372.7478     715.77356  ]
 [1990.6115     485.11075   2105.6335     923.43585  ]
 [2255.5374     568.44867   2326.8682     702.4399   ]
 [ 830.3834     297.209      897.04095    349.59286  ]
 [ 704.9932     419.58734    808.69116    989.6153   ]
 [1646.6818     484.40222   1787.3245     936.4998   ]
 [2437.2583     596.7592    2516.5012     722.8496   ]
 [2176.0432     506.09534   2239.0227     640.1096   ]]

關鍵點預測

pose_results = inference_topdown(pose_estimator, img_path, bboxes) # 元數據
data_samples = merge_data_samples(pose_results) # 把他們打包到一起
# ['pred_instances', '_pred_heatmaps', 'pred_fields', 'gt_instances']

# 每个人 17个关键点 坐标
data_samples.pred_instances.keypoints.shape
(18, 17, 2)
[[[1915.9644    186.88135 ]
  [1927.731     175.11473 ]
  [1892.4312    175.11473 ]
  [1963.0308    198.64795 ]
  [1857.1313    186.88135 ]
  [2010.0973    316.31406 ]
  [1810.0648    328.08066 ]
  [2092.4636    492.8132  ]
  [1786.5316    445.74677 ]
  [1986.564     410.44693 ]
  [1762.9984    504.5798  ]
  [1963.0308    657.5458  ]
  [1821.8314    645.7791  ]
  [1939.4976    939.9444  ]
  [1821.8314    904.64453 ]
  [1927.731    1116.4435  ]
  [1857.1313   1187.0432  ]]

 [[1490.2908    234.2835  ]
  [1514.1499    210.42433 ]
  [1466.4315    210.42433 ]
  [1538.009     210.42433 ]
  [1418.7133    210.42433 ]
  [1585.7273    341.64972 ]

data_samples.gt_instances.bboxes
[[1705.644       96.27843   2114.518     1301.1793   ]
 [1267.5966      94.70741   1701.0554    1316.2964   ]
 [ 720.3899     172.25111   1152.6925    1267.8408   ]
 [   7.2623434  238.24232    170.99622   1140.8058   ]
 [1063.1945     219.22185   1348.985     1244.7227   ]
 [ 436.23294    428.578      586.0933     911.8918   ]
 [ 110.008575   212.1305     324.64334   1123.0087   ]
 [2079.334      478.2407    2224.405      925.2119   ]
 [ 543.71423    406.48438    650.8373     919.45624  ]
 [2363.2996     501.87018   2455.2249     857.3131   ]
 [2308.198      561.452     2372.7478     715.77356  ]
 [1990.6115     485.11075   2105.6335     923.43585  ]
 [2255.5374     568.44867   2326.8682     702.4399   ]
 [ 830.3834     297.209      897.04095    349.59286  ]
 [ 704.9932     419.58734    808.69116    989.6153   ]
 [1646.6818     484.40222   1787.3245     936.4998   ]
 [2437.2583     596.7592    2516.5012     722.8496   ]
 [2176.0432     506.09534   2239.0227     640.1096   ]]

• 可以看每個關鍵點的熱力圖

data_samples.pred_fields.heatmaps.shape
(17, 1418, 2520)

idx_point = 13 # 假設要畫第13個關節點熱力圖
heatmap = data_samples.pred_fields.heatmaps[idx_point,:,:]
plt.imshow(heatmap)
plt.show()

設定半徑跟線寬

pose_estimator.cfg.visualizer.radius = 10
pose_estimator.cfg.visualizer.line_width = 8
visualizer = VISUALIZERS.build(pose_estimator.cfg.visualizer)

visualizer.set_dataset_meta(pose_estimator.dataset_meta) # 一些基本信息，顏色、關節點之間如何連接等等
img = mmcv.imread(img_path)
img = mmcv.imconvert(img, 'bgr', 'rgb')
img_output = visualizer.add_datasample(
            'result',
            img,
            data_sample=data_samples,
            draw_gt=False,
            draw_heatmap=True,
            draw_bbox=True,
            show_kpt_idx=True,
            show=False,
            wait_time=0,
            out_file='outputs/B2.jpg'
)
plt.figure(figsize=(10,10))
plt.imshow(img_output)
plt.show()

訓練自己數據集

建構數據集

• 建議使用pose_tools裡面的工具進行轉換

要進mmpose跟mmdetection進行個別訓練

先訓練目標檢測

• 新增數據集
• 格式
  • mmdetection:
    • data
      • keypoint_coco
        • images
        • train_coco.json
        • val_coco.json
    • rtmdet_m-carpose.py # 我們準備的config檔
  • mmpose:
    • data
      • keypoint_coco
        • images
        • train_coco.json
        • val_coco.json
    • rtmpose-m-carpose.py

訓練時要把numpy版本降下來，因為np.float會出錯(cocoapi)

• 降到1.23.5

mmpose/mmdetection訓練流程

mmdetection

訓練config檔

可見mmdetection/data/rtmdet_m-carpose.py

dataset_type = 'CocoDataset'
data_root = 'data/carpose_keypoint_coco/'
metainfo = {'classes': ('car',)}
NUM_CLASSES = len(metainfo['classes'])

在mmdetection/data/rtmdet_m-carpose.py裡面，可以調整模型大小，可以複製下面模型覆蓋原有模型

# 選擇rtmdet模型
# https://github.com/open-mmlab/mmdetection/tree/3.x/configs/rtmdet
# RTMDet-tiny 
load_from = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet_tiny_8xb32-300e_coco/rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth'
backbone_pretrain = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-tiny_imagenet_600e.pth'
deepen_factor=0.167
widen_factor=0.375
in_channels=[96, 192, 384]
neck_out_channels=96
num_csp_blocks=1
exp_on_reg=False

# RTMDet-s
load_from = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet_s_8xb32-300e_coco/rtmdet_s_8xb32-300e_coco_20220905_161602-387a891e.pth'
backbone_pretrain = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-s_imagenet_600e-ea671761.pth'
deepen_factor=0.33
widen_factor=0.5
in_channels=[128, 256, 512]
neck_out_channels=128
num_csp_blocks=1
exp_on_reg=False

# RTMDet-m
load_from = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet_m_8xb32-300e_coco/rtmdet_m_8xb32-300e_coco_20220719_112220-229f527c.pth'
backbone_pretrain = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-m_8xb256-rsb-a1-600e_in1k-ecb3bbd9.pth'
deepen_factor=0.67
widen_factor=0.75
in_channels=[192, 384, 768]
neck_out_channels=192
num_csp_blocks=2
exp_on_reg=True

# RTMDet-l
load_from = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet_l_8xb32-300e_coco/rtmdet_l_8xb32-300e_coco_20220719_112030-5a0be7c4.pth'
backbone_pretrain = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-l_8xb256-rsb-a1-600e_in1k-6a760974.pth'
deepen_factor=1
widen_factor=1
in_channels=[256, 512, 1024]
neck_out_channels=256
num_csp_blocks=3
exp_on_reg=True

超參數

MAX_EPOCHS = 200
TRAIN_BATCH_SIZE = 4
VAL_BATCH_SIZE = 2
stage2_num_epochs = 20
base_lr = 0.004
VAL_INTERVAL = 5

訓練

cd mmdetection
# 訓練
python tools/train.py data/rtmdet-m-carpose.py
# 測試
python tools/test.py data/rtmdet_m-carpose.py work_dirs/rtmdet_m-carpose/best_coco_bbox_mAP_epoch_190.pth
# 模型權重精簡轉換
python tools/model_converters/publish_model.py work_dirs/rtmdet_m-carpose/best_coco_bbox_mAP_epoch_190.pth checkpoint/rtmdet-m-carpose-200.pth

# 視覺化
cd mmdetection
python vidualize.py # 要去裡面改參數

mmpose

訓練config檔

可見mmpose/data/rtmpose-m-carpose.py

dataset_type = 'CocoDataset'
data_mode = 'topdown'
data_root = 'data/carpose_keypoint_coco/'

• data_root為訓練資料的位置，結構為
  • images
  • train_coco.json
  • val_coco.json

dataset_info = {
    'dataset_name':'carpose_keypoint_coco',
    'classes':'car',
    'paper_info':{
    },
    'keypoint_info':{
        0:{'name':'left-front-bumper','id':0,'color':[0,0,255],'type': '','swap': ''},
        1:{'name':'left-rear-bumper','id':1,'color':[255,0,0],'type': '','swap': ''},
        2:{'name':'left-front-tire','id':2,'color':[255,100,0],'type': '','swap': ''},
        3:{'name':'left-rear-tire','id':3,'color':[0, 255, 0],'type': '','swap': ''},
        4:{'name':'right-front-bumper','id':4,'color':[0, 255, 255],'type': '','swap': ''},
        5:{'name':'right-rear-bumper','id':5,'color':[193, 182, 255],'type': '','swap': ''},
        6:{'name':'right-front-tire','id':6,'color':[193, 100, 255],'type': '','swap': ''},
        7:{'name':'right-rear-tire','id':7,'color':[16, 144, 247],'type': '','swap': ''}
        
    },
    'skeleton_info': {
        0: {'link':('left-front-bumper','left-rear-bumper'),'id': 0,'color': [100,150,200]},
        1: {'link':('left-front-tire','left-rear-tire'),'id': 1,'color': [200,100,150]},
        2: {'link':('right-front-bumper','right-rear-bumper'),'id': 2,'color': [150,120,100]},
        3: {'link':('right-front-bumper','right-rear-bumper'),'id': 3,'color': [150,120,0]}
    }
}

max_epochs = 10 # epoch總數
val_interval = 10 # 每隔幾次保存一次文件
train_cfg = {'max_epochs': max_epochs, 'val_interval': val_interval}
train_batch_size = 32 # 訓練batch
val_batch_size = 8
stage2_num_epochs = 0
base_lr = 4e-3
randomness = dict(seed=21)

# codec settings
codec = dict(
    type='SimCCLabel',
    input_size=(256, 256), # 這邊設定不同尺寸輸入(256、384、512、1024)
    sigma=(12, 12), # 表示轉換高斯分布分布的標準差，越大越容易學習，但精度上限會降低，越小越嚴格，對於人體、人臉高精度場景，可以調小，論文中為5.66
    simcc_split_ratio=2.0,
    normalize=False,
    use_dark=False)

模型

RTMPose-S

model = dict(
    type='TopdownPoseEstimator',
    data_preprocessor=dict(
        type='PoseDataPreprocessor',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        bgr_to_rgb=True),
    backbone=dict(
        _scope_='mmdet',
        type='CSPNeXt',
        arch='P5',
        expand_ratio=0.5,
        deepen_factor=0.33,
        widen_factor=0.5,
        out_indices=(4, ),
        channel_attention=True,
        norm_cfg=dict(type='SyncBN'),
        act_cfg=dict(type='SiLU'),
        init_cfg=dict(
            type='Pretrained',
            prefix='backbone.',
            checkpoint='https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-s_imagenet_600e-ea671761.pth'
        )),
    head=dict(
        type='RTMCCHead',
        in_channels=512,
        out_channels=NUM_KEYPOINTS,
        input_size=codec['input_size'],
        in_featuremap_size=(8, 8),
        simcc_split_ratio=codec['simcc_split_ratio'],
        final_layer_kernel_size=7,
        gau_cfg=dict(
            hidden_dims=256,
            s=128,
            expansion_factor=2,
            dropout_rate=0.,
            drop_path=0.,
            act_fn='SiLU',
            use_rel_bias=False,
            pos_enc=False),
        loss=dict(
            type='KLDiscretLoss',
            use_target_weight=True,
            beta=10.,
            label_softmax=True),
        decoder=codec),
    test_cfg=dict(flip_test=True))

RTMPose-M

model = dict(
    type='TopdownPoseEstimator',
    data_preprocessor=dict(
        type='PoseDataPreprocessor',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        bgr_to_rgb=True),
    backbone=dict(
        _scope_='mmdet',
        type='CSPNeXt',
        arch='P5',
        expand_ratio=0.5,
        deepen_factor=0.67,
        widen_factor=0.75,
        out_indices=(4, ),
        channel_attention=True,
        norm_cfg=dict(type='SyncBN'),
        act_cfg=dict(type='SiLU'),
        init_cfg=dict(
            type='Pretrained',
            prefix='backbone.',
            checkpoint='https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-m_8xb256-rsb-a1-600e_in1k-ecb3bbd9.pth'
        )),
    head=dict(
        type='RTMCCHead',
        in_channels=768,
        out_channels=NUM_KEYPOINTS,
        input_size=codec['input_size'],
        in_featuremap_size=(8, 8),
        simcc_split_ratio=codec['simcc_split_ratio'],
        final_layer_kernel_size=7,
        gau_cfg=dict(
            hidden_dims=256,
            s=128,
            expansion_factor=2,
            dropout_rate=0.,
            drop_path=0.,
            act_fn='SiLU',
            use_rel_bias=False,
            pos_enc=False),
        loss=dict(
            type='KLDiscretLoss',
            use_target_weight=True,
            beta=10.,
            label_softmax=True),
        decoder=codec),
    test_cfg=dict(flip_test=True))

RTMPose-L

model = dict(
    type='TopdownPoseEstimator',
    data_preprocessor=dict(
        type='PoseDataPreprocessor',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        bgr_to_rgb=True),
    backbone=dict(
        _scope_='mmdet',
        type='CSPNeXt',
        arch='P5',
        expand_ratio=0.5,
        deepen_factor=1.,
        widen_factor=1.,
        out_indices=(4, ),
        channel_attention=True,
        norm_cfg=dict(type='SyncBN'),
        act_cfg=dict(type='SiLU'),
        init_cfg=dict(
            type='Pretrained',
            prefix='backbone.',
            checkpoint='https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-l_8xb256-rsb-a1-600e_in1k-6a760974.pth'
        )),
    head=dict(
        type='RTMCCHead',
        in_channels=1024,
        out_channels=NUM_KEYPOINTS,
        input_size=codec['input_size'],
        in_featuremap_size=(8, 8),
        simcc_split_ratio=codec['simcc_split_ratio'],
        final_layer_kernel_size=7,
        gau_cfg=dict(
            hidden_dims=256,
            s=128,
            expansion_factor=2,
            dropout_rate=0.,
            drop_path=0.,
            act_fn='SiLU',
            use_rel_bias=False,
            pos_enc=False),
        loss=dict(
            type='KLDiscretLoss',
            use_target_weight=True,
            beta=10.,
            label_softmax=True),
        decoder=codec),
    test_cfg=dict(flip_test=True))

RTMPose-X

model = dict(
    type='TopdownPoseEstimator',
    data_preprocessor=dict(
        type='PoseDataPreprocessor',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        bgr_to_rgb=True),
    backbone=dict(
        _scope_='mmdet',
        type='CSPNeXt',
        arch='P5',
        expand_ratio=0.5,
        deepen_factor=1.33,
        widen_factor=1.25,
        out_indices=(4, ),
        channel_attention=True,
        norm_cfg=dict(type='SyncBN'),
        act_cfg=dict(type='SiLU'),
        init_cfg=dict(
            type='Pretrained',
            prefix='backbone.',
            checkpoint='https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-x_8xb256-rsb-a1-600e_in1k-b3f78edd.pth'
        )),
    head=dict(
        type='RTMCCHead',
        in_channels=1280,
        out_channels=NUM_KEYPOINTS,
        input_size=codec['input_size'],
        in_featuremap_size=(32, 32),
        simcc_split_ratio=codec['simcc_split_ratio'],
        final_layer_kernel_size=7,
        gau_cfg=dict(
            hidden_dims=256,
            s=128,
            expansion_factor=2,
            dropout_rate=0.,
            drop_path=0.,
            act_fn='SiLU',
            use_rel_bias=False,
            pos_enc=False),
        loss=dict(
            type='KLDiscretLoss',
            use_target_weight=True,
            beta=10.,
            label_softmax=True),
        decoder=codec),
    test_cfg=dict(flip_test=True))

訓練

• 假設上面的config檔檔名為rtmpose-s-carpose.py

cd mmpose
# 訓練
python tools/train.py data/rtmpose-s-carpose.py
# 評估
python tools/test.py data/rtmpose-s-carpose.py \
                      work_dirs/rtmpose-s-carpose/epoch_300.pth
# 模型權重精簡轉換
python tools/misc/publish_model.py \
        work_dirs/rtmpose-s-carpose/epoch_300.pth \
        checkpoint/rtmpose-s-carpose-300.pth

python vidualize.py # 要去裡面改參數