手把手教物体检测——M2Det
目录
在data文件夹下新建VOCdevkit文件夹，导入VOC格式的数据集。如下图：
下载权重文件，放在weights（如果没有就在根目录新建）文件夹下面。
模型介绍
物体检测模型M2Det，是北京大学&阿里达摩院提出的Single-shot目标检测新模型，使用multi-level特征。在MS-COCO benchmark上，M2Det的单尺度版本和多尺度版本AP分别达到41.0和44.2 。
该模型的特点：
- 提出多级特征金字塔网络MLFPN。MLFPN的结构如下：
- 基于提出的MLFPN，结合SSD，提出一种新的Single-shot目标检测模型M2Det
模型使用
-
下载源码:
https://github.com/qijiezhao/M2Det”。
-
在data文件夹下新建VOCdevkit文件夹，导入VOC格式的数据集。如下图：
-
下载权重文件，放在weights（如果没有就在根目录新建）文件夹下面。
-
修改voc0712.py里面的类别。
将：
VOC_CLASSES = ( '__background__', # always index 0
'aeroplane', 'bicycle', 'bird', 'boat',
'bottle', 'bus', 'car', 'cat', 'chair',
'cow', 'diningtable', 'dog', 'horse',
'motorbike', 'person', 'pottedplant',
'sheep', 'sofa', 'train', 'tvmonitor')修改为：
VOC_CLASSES = ( '__background__', # always index 0
'aircraft', 'oiltank')
选择配置文件。
本例采用configs->m2det512_vgg.py配置文件model = dict(
type = 'm2det',
input_size = 512,
init_net = True,
pretrained = 'weights/vgg16_reducedfc.pth',
m2det_config = dict(
backbone = 'vgg16',
net_family = 'vgg', # vgg includes ['vgg16','vgg19'], res includes ['resnetxxx','resnextxxx']
base_out = [22,34], # [22,34] for vgg, [2,4] or [3,4] for res families
planes = 256,
num_levels = 8,
num_scales = 6,
sfam = False,
smooth = True,
num_classes = 3,#更改类别，按照数据集里面的类别数量+1(背景)
),
rgb_means = (104, 117, 123),
p = 0.6,
anchor_config = dict(
step_pattern = [8, 16, 32, 64, 128, 256],
size_pattern = [0.06, 0.15, 0.33, 0.51, 0.69, 0.87, 1.05],
),
save_eposhs = 10,
weights_save = 'weights/' #保存权重文件的目录
)
train_cfg = dict(
cuda = True,#是否使用cuda
warmup = 5,
per_batch_size = 2,#修改batchsize，按照自己显卡的能力修改
lr = [0.004, 0.002, 0.0004, 0.00004, 0.000004],#学利率调整，调整依据step_lr的epoch数值。
gamma = 0.1,
end_lr = 1e-6,
step_lr = dict(
COCO = [90, 110, 130, 150, 160],
VOC = [100, 150, 200, 250, 300], # unsolve
),
print_epochs = 10,#每个10个epoch保存一个模型。
num_workers= 2,#线程数，根据CPU调整
)
test_cfg = dict(
cuda = True,
topk = 0,
iou = 0.45,
soft_nms = True,
score_threshold = 0.1,
keep_per_class = 50,
save_folder = 'eval'
)
loss = dict(overlap_thresh = 0.5,
prior_for_matching = True,
bkg_label = 0,
neg_mining = True,
neg_pos = 3,
neg_overlap = 0.5,
encode_target = False)
optimizer = dict(type='SGD', momentum=0.9, weight_decay=0.0005)#激活函数。
#修改dataset，本例采用VOC2007数据集，将COCO的删除即可,删除VOC2012
dataset = dict(
VOC = dict(
train_sets = [('2007', 'trainval')],
eval_sets = [('2007', 'test')],
)
)
import os
import os
home = ""#home路径，默认是linux的，本例采用win10，讲其修改为“”
VOCroot = os.path.join(home,"data/VOCdevkit/")
COCOroot = os.path.join(home,"data/coco/")-
删除pycocotools
在安装pycocotools工具前提下，将程序自带的pycocotools工具包删除。
修改coco.py
将：
from utils.pycocotools.coco import COCO
from utils.pycocotools.cocoeval import COCOeval
from utils.pycocotools import mask as COCOmask
修改为：
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
from pycocotools import mask as COCOmask
-
修改nms_wrapper.py
将：from .nms.cpu_nms import cpu_nms, cpu_soft_nms
from .nms.gpu_nms import gpu_nms
# def nms(dets, thresh, force_cpu=False):
# """Dispatch to either CPU or GPU NMS implementations."""
# if dets.shape[0] == 0:
# return []
# if cfg.USE_GPU_NMS and not force_cpu:
# return gpu_nms(dets, thresh, device_id=cfg.GPU_ID)
# else:
# return cpu_nms(dets, thresh)
def nms(dets, thresh, force_cpu=False):
"""Dispatch to either CPU or GPU NMS implementations."""
if dets.shape[0] == 0:
return []
if force_cpu:
return cpu_soft_nms(dets, thresh, method = 1)
#return cpu_nms(dets, thresh)
return gpu_nms(dets, thresh)
修改为：from .nms.py_cpu_nms import py_cpu_nms
def nms(dets, thresh, force_cpu=False):
"""Dispatch to either CPU or GPU NMS implementations."""
if dets.shape[0] == 0:
return []
if force_cpu:
return py_cpu_nms(dets, thresh, method = 1)
return py_cpu_nms(dets, thresh)
-
修改train.py
修改选定配置的文件parser.add_argument('-c', '--config', default='configs/m2det512_vgg.py')
修改数据的格式
parser.add_argument('-d', '--dataset', default='VOC', help='VOC or COCO dataset')
然后就可以开始训练了。-
修改test.py
parser = argparse.ArgumentParser(description='M2Det Testing')
parser.add_argument('-c', '--config', default='configs/m2det512_vgg.py', type=str)#选择配置文件，和训练的配置文件对应
parser.add_argument('-d', '--dataset', default='VOC', help='VOC or COCO version')
parser.add_argument('-m', '--trained_model', default='weights/M2Det_VOC_size512_netvgg16_epoch30.pth', type=str, help='Trained state_dict file path to open')
parser.add_argument('--test', action='store_true', help='to submit a test file')
修改voc0712.py282行的xml路径。将：annopath = os.path.join(
rootpath,
'Annotations',
'{:s}.xml')
改为：annopath = rootpath+'/Annotations/{:s}.xml'
测试结果：
-
可视化结果
修改demo.py中超参数
parser.add_argument('-c', '--config', default='configs/m2det512_vgg.py', type=str)
parser.add_argument('-f', '--directory', default='imgs/', help='the path to demo images')
parser.add_argument('-m', '--trained_model', default='weights/M2Det_VOC_size512_netvgg16_epoch30.pth', type=str, help='Trained state_dict file path to open')
然后将部分测试图片放到imgs文件夹下面，运行demo.py.
本文实例：https://download.csdn.net/download/hhhhhhhhhhwwwwwwwwww/12428492