FastestDet: it has higher accuracy and faster speed than Yolo-fastest https://github.com/dog-qiuqiu/FastestDet
根据github上的一个Yolo-FastestV2参考作为参考，以及gd32ai-modelzoo的代码进行复现，选用TT100k交通标志检测数据集完成检测与分类的任务。模型需要部署到开发板上，对推理速度有一定的要求，最后选用YoloFastestV2模型。
因为我们需要将模型部署到gd32开发板上，而摄像头采用ov7670采集图像最大为640*480,而数据集的图片大小为2048*2048，所以我们将数据集的图像需要重新做切割。我按照每个图片的每一个标签的中心点做切割，切割成320*240大小的图像，这样将多目标的数据集转换成了单目标检测的数据集。再将标签数据归一化，转换成了yolo所需的txt文件标签。
# -*- coding: UTF-8 -*-
# @Project ：group_first
# @File ：preprocessing.py
# @IDE ：PyCharm
# @Author ：[email protected]
# @Date ：2024/5/7 15:55
import json
import random
from PIL import ImageDraw, Image
import os
data_path = './data/annotations.json'
train_path = './data/train'
test_path = './data/test'
def save_json(save_path,data):
assert save_path.split('.')[-1] == 'json'
with open(save_path,'w') as file:
json.dump(data,file)
def convert(size, box):
dw = 1. / (size[0])
dh = 1. / (size[1])
x = box[0] + box[2] / 2.0
y = box[1] + box[3] / 2.0
w = box[2]
h = box[3]
x = x * dw
w = w * dw
y = y * dh
h = h * dh
return (x, y, w, h)
def print_progress_bar(current, total, prefix='', suffix='', decimals=1, length=50, fill='█', print_end="\r"):
"""
调用时需要提供当前进度和总进度。
@params:
current - 当前的进度值（Int）
total - 总进度值（Int）
prefix - 前缀字符串（Str）
suffix - 后缀字符串（Str）
decimals - 百分比精度（Int）
length - 进度条长度（Int）
fill - 填充字符（Str）
print_end - 结束打印字符（Str）
"""
percent = ("{0:." + str(decimals) + "f}").format(100 * (current / float(total)))
filled_length = int(length * current // total)
bar = fill * filled_length + '-' * (length - filled_length)
print(f'\r{prefix} |{bar}| {percent}% {suffix}', end=print_end)
# 当进度条完成时换行
if current == total:
print()
os.system('cls')
def normalize_bbox(bbox, image_width, image_height):
"""
将边界框归一化到 [0, 1] 的范围。
参数:
bbox (list or tuple): 边界框的坐标，格式为 [x_min, y_min, x_max, y_max]。
image_width (int or float): 图像的宽度。
image_height (int or float): 图像的高度。
返回:
list: 归一化后的边界框坐标。
"""
x_min, y_min, x_max, y_max = bbox
x_min_normalized = x_min / image_width
y_min_normalized = y_min / image_height
x_max_normalized = x_max / image_width
y_max_normalized = y_max / image_height
return [x_min_normalized, y_min_normalized, x_max_normalized, y_max_normalized]
# 获取训练图片图片id
trains_id_ = os.listdir(train_path)
trains_id = [] # 图像id
for train_id in trains_id_:
trains_id.append(train_id.split('.')[0])
# 获取测试图片图片id
tests_id_ = os.listdir(test_path)
tests_id = [] # 图像id
for test_id in tests_id_:
tests_id.append(test_id.split('.')[0])
print('训练ID：',trains_id)
print('测试ID：',tests_id)
# 获取标签
f = open(data_path, 'r')
content = f.read()
annotations = json.loads(content)
# lable_name = []
with open('./dataset/trc.txt','r',encoding='utf-8') as f:
lable_name = [line.strip() for line in f.readlines()]
for key,value in enumerate(lable_name):
print(key,value)
width, height = 320,240
owidth,oheight = 2048,2048
def preprocessing(s_id,is_train):
if is_train is True:
open_path = train_path
save_path = "./dataset/train"
else:
open_path = test_path
save_path = "./dataset/val"
sum_id = s_id
for id in s_id:
sum_id.pop(0)
print_progress_bar(len(sum_id), len(s_id), prefix='Progress:', suffix='Complete', length=50)
# print(f'处理：{id}中。。。')
target_dir = annotations['imgs'][f'{id}']
path = target_dir['path']
path = path.split('.')[0]
object = target_dir['objects']
"""
resize图像
"""
is_save = False
encode = 0
for obj in object:
encode +=1
# if object[0]['category'] not in lable_name:
# lable_name.append(object[0]['category'])
xmin = obj['bbox']['xmin']
ymin = obj['bbox']['ymin']
# 右下点
ymax = obj['bbox']['ymax']
xmax = obj['bbox']['xmax']
# print('原坐标：',xmin, ymin, xmax, ymax)
# 计算中心点
x_center = (xmin + xmax) / 2 + random.randint(-100, 100)
y_center = (ymin + ymax) / 2 + random.randint(-100, 100)
# print('原标签框位置：',xmin,ymax,xmax,ymin)
# 计算图像边缘
img_xmin = x_center - 160
img_xmax = x_center + 160
img_ymin = y_center - 120
img_ymax = y_center + 120
new_xmin = abs(xmin -img_xmin)
new_ymin = abs(ymin -img_ymin)
new_xmax = abs(new_xmin + (xmax - xmin))
new_ymax = abs(new_ymin + (ymax - ymin))
try:
index = lable_name.index(obj['category'])
boxx = [new_xmin, new_ymin, new_xmax, new_ymax]
size = (width, height)
# x, y, w, h = convert(size, boxx)
x, y, w, h = normalize_bbox(boxx, size[0], size[1])
# print('归一化：',x, y, w, h)
xmin = str(x)
ymin = str(y)
xmax = str(w)
ymax = str(h)
content = str(index + 1) + ' ' + xmin + ' ' + ymin + ' ' + xmax + ' ' + ymax
with open(f"{save_path}/{id}_{encode}.txt", 'a') as f:
f.write(content + '\n')
except ValueError:
print(f"{id}编号的标签为{obj['category']}，列表中没有此标签")
continue
try:
img = Image.open(os.path.join(open_path, f'{id}.jpg'))
# cropped_img.show()
cropped_img = img.crop((img_xmin, img_ymin, img_xmax, img_ymax))
# cropped_img.show()
# cropped_img.save(f'./dataset/{open_path}/{train_id}.jpg')
cropped_img.save(os.path.join(save_path, f"{id}_{encode}.jpg"))
is_save = True
if is_save is True:
# todo：写入文件路径
if is_train is True:
with open(f'./dataset/train.txt', 'a', encoding='utf-8') as f:
f.write(path+ f"_{encode}.jpg" + '\n')
else:
with open(f'./dataset/val.txt', 'a', encoding='utf-8') as f:
f.write(path+ f"_{encode}.jpg" + '\n')
except FileNotFoundError:
print(f'{open_path}\{id}.jpg 不存在')
if __name__ == '__main__':
for id,is_train in (trains_id,True),(tests_id,False):
preprocessing(id,is_train)
# print(len(lable_name))
# with open(".\dataset\data.txt", 'w') as f:
# for i in lable_name:
# f.write(i + '\n')能力有限，并未对Yolo-FastestV2做很大的改动，只是对该项目在本机中运行做遇到问题，做了一点改动。代码也上传至我的github。
这样的结果其实我也很惊讶，即使是单目标检测，结果精度都这么低，相比较于coco数据集精度差了将近一倍，我暂时觉得是因为目标太小的问题。我暂时无法解决，我看到Yolo-FastestV2的作者有采用FPN模型，但是精度还是很低，其实还可以尝试采用其他模型作为训练，但我毕竟不是CV方向，并且时间不够，就没有继续往下钻研。我觉得目标小，有点类似于遥感的目标检测，可以采用遥感的目标检测算法，来检测该数据集（如过不是在嵌入式单片机上跑的话）可能会有好的结果出来。
D:\anaconda\envs\gd32ai\lib\site-packages\torch\functional.py:512: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at C:\cb\pytorch_1000000000000\work\aten\src\ATen\native\TensorShape.cpp:3588.)
return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
Evaluation model:: 100%|██████████| 60/60 [00:38<00:00, 1.54it/s]
computer PR...
Evaluation model:: 100%|██████████| 60/60 [00:10<00:00, 5.74it/s]
Precision:0.281061 Recall:0.445965 AP:0.318521 F1:0.337140
- Simple, fast, compact, easy to transplant
- Less resource occupation, excellent single-core performance, lower power consumption
- Faster and smaller:Trade 0.3% loss of accuracy for 30% increase in inference speed, reducing the amount of parameters by 25%
- Fast training speed, low computing power requirements, training only requires 3GB video memory, gtx1660ti training COCO 1 epoch only takes 4 minutes
- 算法介绍：https://zhuanlan.zhihu.com/p/400474142 交流qq群:1062122604
| Network | COCO mAP(0.5) | Resolution | Run Time(4xCore) | Run Time(1xCore) | FLOPs(G) | Params(M) |
|---|---|---|---|---|---|---|
| Yolo-FastestV2 | 24.10 % | 352X352 | 3.29 ms | 5.37 ms | 0.212 | 0.25M |
| Yolo-FastestV1.1 | 24.40 % | 320X320 | 4.23 ms | 7.54 ms | 0.252 | 0.35M |
| Yolov4-Tiny | 40.2% | 416X416 | 26.00ms | 55.44ms | 6.9 | 5.77M |
- Test platform Mate 30 Kirin 990 CPU，Based on NCNN
- Different loss weights for different scale output layers
- The backbone is replaced with a more lightweight shufflenetV2
- Anchor matching mechanism and loss are replaced by YoloV5, and the classification loss is replaced by softmax cross entropy from sigmoid
- Decouple the detection head, distinguish obj (foreground background classification), cls (category classification), reg (detection frame regression) 3 branches,
- PIP
pip3 install -r requirements.txt
- Picture test
python3 test.py --data data/coco.data --weights modelzoo/coco2017-0.241078ap-model.pth --img img/000139.jpg
/>
-
The format of the data set is the same as that of Darknet Yolo, Each image corresponds to a .txt label file. The label format is also based on Darknet Yolo's data set label format: "category cx cy wh", where category is the category subscript, cx, cy are the coordinates of the center point of the normalized label box, and w, h are the normalized label box The width and height, .txt label file content example as follows:
11 0.344192634561 0.611 0.416430594901 0.262 14 0.509915014164 0.51 0.974504249292 0.972 -
The image and its corresponding label file have the same name and are stored in the same directory. The data file structure is as follows:
. ├── train │ ├── 000001.jpg │ ├── 000001.txt │ ├── 000002.jpg │ ├── 000002.txt │ ├── 000003.jpg │ └── 000003.txt └── val ├── 000043.jpg ├── 000043.txt ├── 000057.jpg ├── 000057.txt ├── 000070.jpg └── 000070.txt -
Generate a dataset path .txt file, the example content is as follows：
train.txt
/home/qiuqiu/Desktop/dataset/train/000001.jpg /home/qiuqiu/Desktop/dataset/train/000002.jpg /home/qiuqiu/Desktop/dataset/train/000003.jpgval.txt
/home/qiuqiu/Desktop/dataset/val/000070.jpg /home/qiuqiu/Desktop/dataset/val/000043.jpg /home/qiuqiu/Desktop/dataset/val/000057.jpg -
Generate the .names category label file, the sample content is as follows:
category.names
person bicycle car motorbike ... -
The directory structure of the finally constructed training data set is as follows:
. ├── category.names # .names category label file ├── train # train dataset │ ├── 000001.jpg │ ├── 000001.txt │ ├── 000002.jpg │ ├── 000002.txt │ ├── 000003.jpg │ └── 000003.txt ├── train.txt # train dataset path .txt file ├── val # val dataset │ ├── 000043.jpg │ ├── 000043.txt │ ├── 000057.jpg │ ├── 000057.txt │ ├── 000070.jpg │ └── 000070.txt └── val.txt # val dataset path .txt file
- Generate anchor based on current dataset
python3 genanchors.py --traintxt ./train.txt - The anchors6.txt file will be generated in the current directory,the sample content of the anchors6.txt is as follows:
12.64,19.39, 37.88,51.48, 55.71,138.31, 126.91,78.23, 131.57,214.55, 279.92,258.87 # anchor bias 0.636158 # iou
- Reference./data/coco.data
[name] model_name=coco # model name [train-configure] epochs=300 # train epichs steps=150,250 # Declining learning rate steps batch_size=64 # batch size subdivisions=1 # Same as the subdivisions of the darknet cfg file learning_rate=0.001 # learning rate [model-configure] pre_weights=None # The path to load the model, if it is none, then restart the training classes=80 # Number of detection categories width=352 # The width of the model input image height=352 # The height of the model input image anchor_num=3 # anchor num anchors=12.64,19.39, 37.88,51.48, 55.71,138.31, 126.91,78.23, 131.57,214.55, 279.92,258.87 #anchor bias [data-configure] train=/media/qiuqiu/D/coco/train2017.txt # train dataset path .txt file val=/media/qiuqiu/D/coco/val2017.txt # val dataset path .txt file names=./data/coco.names # .names category label file
- Perform training tasks
python3 train.py --data data/coco.data
- Calculate map evaluation
python3 evaluation.py --data data/coco.data --weights modelzoo/coco2017-0.241078ap-model.pth
- Convert onnx
python3 pytorch2onnx.py --data data/coco.data --weights modelzoo/coco2017-0.241078ap-model.pth --output yolo-fastestv2.onnx - onnx-sim
python3 -m onnxsim yolo-fastestv2.onnx yolo-fastestv2-opt.onnx - Build NCNN
git clone https://github.com/Tencent/ncnn.git cd ncnn mkdir build cd build cmake .. make make install cp -rf ./ncnn/build/install/* ~/Yolo-FastestV2/sample/ncnn - Covert ncnn param and bin
cd ncnn/build/tools/onnx ./onnx2ncnn yolo-fastestv2-opt.onnx yolo-fastestv2.param yolo-fastestv2.bin cp yolo-fastestv2* ../ cd ../ ./ncnnoptimize yolo-fastestv2.param yolo-fastestv2.bin yolo-fastestv2-opt.param yolo-fastestv2-opt.bin 1 cp yolo-fastestv2-opt* ~/Yolo-FastestV2/sample/ncnn/model - run sample
cd ~/Yolo-FastestV2/sample/ncnn sh build.sh ./demo