dzh358/zisan

Integrated CV basic application development kit based on Pytorch 基于Pytorch的集成CV基础应用开发工具包

Install zisan

JintuZheng Jan 25th 2020

Last page: wiki: Strat pytorch

Step1: Install zisan package

Two easy ways to install zisan. Input the follow command:

pip install zisan

Also you can download the wheel file on the official site: http://jintupersonal.com/zisan/

Pay attention to the lastest version to modify your pip command.

Step2: Prepare the weight files for zisan

Firstly you should download the weights files for zisan. You can find the url button on homepage: Download weights link：https://pan.baidu.com/s/1qj-Lpe4OKV0L-w9uKO8EFw key：x9wl The weight files structure is followed:

Instance_Seg_weight (Folder)

Jintu_SEG_Interactive.pth (178 MB)

ObjectDetect_data_weights (Folder)

runBox.zip (475 MB)

Unzip runBox.zip, you will see three folders, they are "cfgs", "data", "weights".

cfgs: Here put the network configs

data: Put your data preparing to train

weights: Here save the raw yolo weights and your last output weights

IMPORTANT !

DON'T RECHANGE THE FOLDER'S NAME !

DON'T RECHANGE THE FOLDER'S NAME !

DON'T RECHANGE THE FOLDER'S NAME !

you can't rechange the "cfgs", "weights", "data" and their child folders name, because zisan. Objedetect class will use the refer all folders' name, if you rechange them, it will happen errors. But you can set your current path and as parameters in function.

more details you can refer the following page: wiki: Train your dataset

How to use these weights? You will find your answer in the following Demo Courses.

Next page: wiki: Class ImgSeg

Demo: A person segmentation

JintuZheng Jan 26th 2020

Raw page: wiki: Demo: A person segmentation Last page: wiki: Demo: A Box segmentation

Prepare a person picture

(This picture is from dataset davis2017) Find the person and give him bone marks, like the following. Maybe it's very abstract. Monofilament doesn't affect our segmentation of objects。

Import Packages

from zisan.Seg.Interface import ImgSeg, markTools
import os
import numpy as np
import cv2
from PIL import Image

Draw mask

lines=[[(281,120),(267,341)],[(279,157),(208,171)],[(309,170),(308,250)],[(275,233),(370,341)]]
img=Image.open(current_path+'/temp/1.jpg').convert('RGB')
markpen=markTools(img.height,img.width)
for line in lines:
	markpen.curveDraw(line,is_Pos=True)

[(281,120),(267,341)] is a line element lines is a line list=[[(,),(,)]]

Instance segmentation

re=markpen.getMark_result(is_showPreview=True)
model.ImgSeg_SingleObj(img,re,is_showPreview=True)

Binary mask result preview:

Fade mask show preview:

zisan have the Yolov3 interface to train your own dataset. Please view the next part: wiki: Package: Object detect