caffe train face licenseplate reID action ocr

---

2020.07.11. Updates: optimizing the CUDA memory to improve the batch_size in one batch at the limited Cuda memory of 16GitaBytes.

---

caffe implement method

Method	orginal performance	caffe performance
focal loss	NOT TEST	NOT TEST BUT UESD IN THE CENTERNET
data_anchor_sample	SSD + DAS +VOC Paddlepaddle 72.2%	SSD + DAS +VOC caffe 74.1%
cos_loss	NOT TESTED	NOT TESTED CAFFE
Yolov3	Yolov3+darknet COCO 57.9%	Yolov3 Caffe COCO 53.2%
triplet_loss	facenet+tf 99.85%	mobilenet-face 99.4%

---

centernet face + nms version

Method	Easy	Medium	Hard
ours(one scale(640x640) + nms)	0.8388	0.8223	0.7246
ours(one scale(640x640))	0.8326	0.8147	0.6955
ours(one scale(800x800) + nms)	0.850	0.863	0.846
original	0.922	0.911	0.782

---

example val accuray

Method	accuray performance
mobilenet-v2 face recognition	99.42%
face landmarks + face attributes gender	(99.4%)+ bool glasses(99.5%)
face head angle	(not evaluated)
car license plate detect	(mAP91%)
car recognition	(96%, only support anhui car + blue car style)

New proposed method CenterGridSoftmax + nms method by using anchor

Method	Easy	Medium	Hard
ours(one scale)640x640 + nms	0.8636	0.8489	0.7425

---

will add caffe version efficientDet, to be continued

we had add the net prototxt, not training

python Caffe API

import caffe
from caffe import layers as L
from caffe import params as P
L.Data( source=lmdb,backend=P.Data.LMDB,batch_size=batch_size, ntop=2,transform_param=dict(crop_size=227,mean_value=[104, 117, 123],mirror=True))
L.HDF5Data(hdf5_data_param={'source': './training_data_paths.txt','batch_size': 64},include={'phase': caffe.TRAIN})
L.ImageData(source=list_path,batch_size=batch_size,new_width=48,new_height=48,ntop=2,ransform_param=dict(crop_size=40,mirror=True))
L.Convolution(bottom, kernel_size=ks, stride=stride,num_output=nout, pad=pad, group=group)
L.LRN(bottom,local_size=5, alpha=1e-4, beta=0.75)
L.ReLU(bottom, in_place=True)
L.Pooling(bottom, pool=P.Pooling.MAX, kernel_size=ks, stride=stride)
L.InnerProduct(bottom, num_output=nout)
L.Dropout(bottom, in_place=True)
L.SoftmaxWithLoss(bottom, label)

查看某一层的宽，高，shape，可以用net['layer_name'].blobs[].data.shape[]

---

数据增强方法

假设原图输入是一张640*480的图片，这里由于版面问题我放缩了图片尺寸并且没做mean subtract，由于最后会有resize参数导致输出的图片都会resize到300x300，但是主要看的是增强的效果，SSD中的数据增强的顺序是：

DistortImage: 这个主要是修改图片的brightness，contrast，saturation，hue，reordering channels，并没改变标签bbox

ExpandImage: 这个主要是将DistortImage的图片用像素0进行扩展，标签bbox此时肯定会改变，就重新以黑边的左上角为原点计算[0,1]的bbox的左上角和右下角两个点坐标。

BatchSampler: 由于这里选错图了，BatchSampler必须要有GT的存在才会生效，由于我做的是人的检测所以图中没人就不会生成sampled_bboxes，后面修改例子。sampled_bboxes的值是随机在[0, 1]上生成的bbox，并且和某个gt_bboxes的IOU在[min, max]之间。由于proto中配的max_sample都是为1，所以每个batch_sampler可能会有1个sampled_bbox，随机取一个sampled bbox并且裁剪图片和标签。标签裁剪也很好理解首先要通过ProjectBBox将原坐标系标签投影到裁剪后图片的新坐标系的坐标，然后再ClipBBox到[0,1]之间。

Resize：放缩到300x300，最后将图片放缩到300x300，标签框也是线性放缩坐标而已。

Crop：原本data_transformer还会crop的，这个参数是配在prototxt中，默认是原图 所以就和没crop一样。如果要crop的话标签也是会和之前BatchSampler那样处理。

data_anchor_sampler: 在一张图像中随机抽取一张sface