lichen/detecttracking
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

last update in 2024

Browse Source
This commit is contained in:
王庆刚
2024-12-31 16:45:04 +08:00
parent dac3b3f2b6
commit 7e13e0f5b4
20 changed files with 1349 additions and 389 deletions
Download Patch File Download Diff File Expand all files Collapse all files

291
contrast/one2one_contrast.py
View File

@ -11,7 +11,7 @@ Created on Fri Aug 30 17:53:03 2024
标准特征提取，并保存至文件夹 stdFeaturePath 中，
也可在运行过程中根据与购物事件集合 barcodes 交集执行
2. 1:1 比对性能测试，
func: one2one_eval(similPath)
func: one2one_simi()
(1) 求购物事件和标准特征级 Barcode 交集，构造 evtDict、stdDict
(2) 构造扫 A 放 A、扫 A 放 B 组合，mergePairs = AA_list + AB_list
(3) 循环计算 mergePairs 中元素 "(A, A) 或 (A, B)" 相似度;
@ -20,7 +20,7 @@ Created on Fri Aug 30 17:53:03 2024
3. precise、recall等指标计算
func: compute_precise_recall(pickpath)
func: compute_one2one_pr(pickpath)
@author: ym
@ -33,6 +33,9 @@ import sys
import random
import pickle
import json
import random
import copy
import sys
# import torch
import time
# import json
@ -54,6 +57,7 @@ from feat_extract.config import config as conf
from feat_extract.inference import FeatsInterface
from utils.event import ShoppingEvent, save_data
from genfeats import gen_bcd_features
from event_test import calc_simil
@ -175,14 +179,53 @@ def data_precision_compare(stdfeat, evtfeat, evtMessage, save=True):
f.write(line + '\n')
def one2one_simi():
def simi_calc(event, stdfeat):
evtfeat = event.feats_compose
if isinstance(event.feats_select, list):
if len(event.feats_select) and len(event.feats_select[0]):
evtfeat = event.feats_select[0]
else:
return None, None, None
else:
evtfeat = event.feats_select
if len(evtfeat)==0 or len(stdfeat)==0:
return None, None, None
matrix = 1 - cdist(evtfeat, stdfeat, 'cosine')
matrix[matrix < 0] = 0
simi_mean = np.mean(matrix)
simi_max = np.max(matrix)
stdfeatm = np.mean(stdfeat, axis=0, keepdims=True)
evtfeatm = np.mean(evtfeat, axis=0, keepdims=True)
simi_mfeat = 1- np.maximum(0.0, cdist(stdfeatm, evtfeatm, 'cosine'))
return simi_mean, simi_max, simi_mfeat[0,0]
def build_std_evt_dict():
'''
stdFeaturePath: 标准特征集地址
eventDataPath: Event对象地址
'''
stdBarcode = [p.stem for p in Path(stdFeaturePath).iterdir() if p.is_file() and p.suffix=='.pickle']
# stdBarcode = [p.stem for p in Path(stdFeaturePath).iterdir() if p.is_file() and p.suffix=='.json']
'''*********** USearch ***********'''
stdFeaturePath = r"D:\contrast\stdlib\v11_test.json"
stdBarcode = []
stdlib = {}
with open(stdFeaturePath, 'r', encoding='utf-8') as f:
data = json.load(f)
for dic in data['total']:
barcode = dic['key']
feature = np.array(dic['value'])
stdBarcode.append(barcode)
stdlib[barcode] = feature
'''======1. 购物事件列表，该列表中的 Barcode 存在于标准的 stdBarcode 内 ==='''
evtList = [(p.stem, p.stem.split('_')[-1]) for p in Path(eventDataPath).iterdir()
if p.is_file()
@ -192,16 +235,21 @@ def one2one_simi():
and p.stem.split('_')[-1] in stdBarcode
]
barcodes = set([bcd for _, bcd in evtList])
'''======2. 构建用于比对的标准特征字典 ============='''
# stdDict = {}
# for barcode in barcodes:
# stdpath = os.path.join(stdFeaturePath, barcode+'.json')
# with open(stdpath, 'r', encoding='utf-8') as f:
# stddata = json.load(f)
# feat = np.array(stddata["value"])
# stdDict[barcode] = feat
'''*********** USearch ***********'''
stdDict = {}
for barcode in barcodes:
stdpath = os.path.join(stdFeaturePath, barcode+'.pickle')
with open(stdpath, 'rb') as f:
stddata = pickle.load(f)
stdDict[barcode] = stddata
stdDict[barcode] = stdlib[barcode]
'''======3. 构建用于比对的操作事件字典 ============='''
evtDict = {}
for evtname, barcode in evtList:
@ -209,21 +257,123 @@ def one2one_simi():
with open(evtpath, 'rb') as f:
evtdata = pickle.load(f)
evtDict[evtname] = evtdata
return evtList, evtDict, stdDict
def one2SN_pr(evtList, evtDict, stdDict):
std_barcodes = set([bcd for _, bcd in evtList])
tp_events, fn_events, fp_events, tn_events = [], [], [], []
tp_simi, fn_simi, tn_simi, fp_simi = [], [], [], []
errorFile_one2SN = []
SN = 9
for evtname, barcode in evtList:
bcd_selected = [barcode]
dset = list(std_barcodes - set([barcode]))
if len(dset) > SN:
random.shuffle(dset)
bcd_selected.extend(dset[:SN])
else:
bcd_selected.extend(dset)
event = evtDict[evtname]
## 无轨迹判断
if len(event.front_feats)+len(event.back_feats)==0:
print(evtname)
continue
barcodes, similars = [], []
for stdbcd in bcd_selected:
stdfeat = stdDict[stdbcd]
# simi_mean, simi_max, simi_mfeat = simi_calc(event, stdfeat)
simi_mean = calc_simil(event, stdfeat)
## 在event.front_feats和event.back_feats同时为空时，此处不需要保护
# if simi_mean==None:
# continue
barcodes.append(stdbcd)
similars.append(simi_mean)
## 此处不需要保护
# if len(similars)==0:
# print(evtname)
# continue
max_idx = similars.index(max(similars))
max_sim = similars[max_idx]
for i in range(len(barcodes)):
bcd, simi = barcodes[i], similars[i]
if bcd==barcode and simi==max_sim:
tp_simi.append(simi)
tp_events.append(evtname)
elif bcd==barcode and simi!=max_sim:
fn_simi.append(simi)
fn_events.append(evtname)
elif bcd!=barcode and simi!=max_sim:
tn_simi.append(simi)
tn_events.append(evtname)
elif bcd!=barcode and simi==max_sim and barcode in barcodes:
fp_simi.append(simi)
fp_events.append(evtname)
else:
errorFile_one2SN.append(evtname)
PPreciseX, PRecallX = [], []
NPreciseX, NRecallX = [], []
Thresh = np.linspace(-0.2, 1, 100)
for th in Thresh:
'''适用于 (Precise, Recall) 计算方式：多个相似度计算并排序，barcode相等且排名第一为 TP '''
'''===================================== 1:SN '''
TPX = sum(np.array(tp_simi) >= th)
FPX = sum(np.array(fp_simi) >= th)
FNX = sum(np.array(fn_simi) < th)
TNX = sum(np.array(tn_simi) < th)
PPreciseX.append(TPX/(TPX+FPX+1e-6))
PRecallX.append(TPX/(len(tp_simi)+len(fn_simi)+1e-6))
NPreciseX.append(TNX/(TNX+FNX+1e-6))
NRecallX.append(TNX/(len(tn_simi)+len(fp_simi)+1e-6))
fig, ax = plt.subplots()
ax.plot(Thresh, PPreciseX, 'r', label='Precise_Pos: TP/TPFP')
ax.plot(Thresh, PRecallX, 'b', label='Recall_Pos: TP/TPFN')
ax.plot(Thresh, NPreciseX, 'g', label='Precise_Neg: TN/TNFP')
ax.plot(Thresh, NRecallX, 'c', label='Recall_Neg: TN/TNFN')
ax.set_xlim([0, 1])
ax.set_ylim([0, 1])
ax.grid(True)
ax.set_title('1:SN Precise & Recall')
ax.set_xlabel(f"Event Num: {len(evtList)}")
ax.legend()
plt.show()
## ============================= 1:N 展厅 直方图'''
fig, axes = plt.subplots(2, 2)
axes[0, 0].hist(tp_simi, bins=60, range=(-0.2, 1), edgecolor='black')
axes[0, 0].set_xlim([-0.2, 1])
axes[0, 0].set_title(f'TP({len(tp_simi)})')
axes[0, 1].hist(fp_simi, bins=60, range=(-0.2, 1), edgecolor='black')
axes[0, 1].set_xlim([-0.2, 1])
axes[0, 1].set_title(f'FP({len(fp_simi)})')
axes[1, 0].hist(tn_simi, bins=60, range=(-0.2, 1), edgecolor='black')
axes[1, 0].set_xlim([-0.2, 1])
axes[1, 0].set_title(f'TN({len(tn_simi)})')
axes[1, 1].hist(fn_simi, bins=60, range=(-0.2, 1), edgecolor='black')
axes[1, 1].set_xlim([-0.2, 1])
axes[1, 1].set_title(f'FN({len(fn_simi)})')
plt.show()
'''======4.2 barcode 标准图像保存 =================='''
# for stdbcd in barcodes:
# stdImgpath = stdDict[stdbcd]["imgpaths"]
# pstdpath = os.path.join(subimgPath, f"{stdbcd}")
# if not os.path.exists(pstdpath):
# os.makedirs(pstdpath)
# ii = 1
# for filepath in stdImgpath:
# stdpath = os.path.join(pstdpath, f"{stdbcd}_{ii}.png")
# shutil.copy2(filepath, stdpath)
# ii += 1
'''======5 构造 3 个事件对: 扫 A 放 A, 扫 A 放 B, 合并 ===================='''
def one2one_simi(evtList, evtDict, stdDict):
barcodes = set([bcd for _, bcd in evtList])
'''======1 构造 3 个事件对: 扫 A 放 A, 扫 A 放 B, 合并 ===================='''
AA_list = [(evtname, barcode, "same") for evtname, barcode in evtList]
AB_list = []
for evtname, barcode in evtList:
@ -234,45 +384,36 @@ def one2one_simi():
mergePairs = AA_list + AB_list
'''======6 计算事件、标准特征集相似度 =================='''
'''======2 计算事件、标准特征集相似度 =================='''
rltdata = []
for i in range(len(mergePairs)):
evtname, stdbcd, label = mergePairs[i]
event = evtDict[evtname]
##============================================ float32
stdfeat = stdDict[stdbcd]["feats_ft32"]
evtfeat = event.feats_compose
if len(evtfeat)==0: continue
matrix = 1 - cdist(stdfeat, evtfeat, 'cosine')
matrix[matrix < 0] = 0
simi_mean = np.mean(matrix)
simi_max = np.max(matrix)
stdfeatm = np.mean(stdfeat, axis=0, keepdims=True)
evtfeatm = np.mean(evtfeat, axis=0, keepdims=True)
simi_mfeat = 1- np.maximum(0.0, cdist(stdfeatm, evtfeatm, 'cosine'))
rltdata.append((label, stdbcd, evtname, simi_mean, simi_max, simi_mfeat[0,0]))
if len(event.feats_compose)==0: continue
stdfeat = stdDict[stdbcd] # float32
simi_mean, simi_max, simi_mfeat = simi_calc(event, stdfeat)
if simi_mean is None:
continue
rltdata.append((label, stdbcd, evtname, simi_mean, simi_max, simi_mfeat))
'''================ float32、16、int8 精度比较与存储 ============='''
# data_precision_compare(stdfeat, evtfeat, mergePairs[i], save=True)
print("func: one2one_eval(), have finished!")
return rltdata
def compute_precise_recall(rltdata):
def one2one_pr(rltdata):
Same, Cross = [], []
for label, stdbcd, evtname, simi_mean, simi_max, simi_mft in rltdata:
if label == "same":
Same.append(simi_mean)
Same.append(simi_max)
if label == "diff":
Cross.append(simi_mean)
Cross.append(simi_max)
Same = np.array(Same)
Cross = np.array(Cross)
@ -280,11 +421,11 @@ def compute_precise_recall(rltdata):
TNFP = len(Cross)
# fig, axs = plt.subplots(2, 1)
# axs[0].hist(Same, bins=60, edgecolor='black')
# axs[0].hist(Same, bins=60, range=(-0.2, 1), edgecolor='black')
# axs[0].set_xlim([-0.2, 1])
# axs[0].set_title(f'Same Barcode, Num: {TPFN}')
# axs[1].hist(Cross, bins=60, edgecolor='black')
# axs[1].hist(Cross, bins=60, range=(-0.2, 1), edgecolor='black')
# axs[1].set_xlim([-0.2, 1])
# axs[1].set_title(f'Cross Barcode, Num: {TNFP}')
# plt.savefig(f'./result/{file}_hist.png') # svg, png, pdf
@ -324,6 +465,23 @@ def compute_precise_recall(rltdata):
rltpath = os.path.join(similPath, 'pr.png')
plt.savefig(rltpath) # svg, png, pdf
fig, axes = plt.subplots(2,1)
axes[0].hist(Same, bins=60, range=(-0.2, 1), edgecolor='black')
axes[0].set_xlim([-0.2, 1])
axes[0].set_title(f'TP({len(Same)})')
axes[1].hist(Cross, bins=60, range=(-0.2, 1), edgecolor='black')
axes[1].set_xlim([-0.2, 1])
axes[1].set_title(f'TN({len(Cross)})')
rltpath = os.path.join(similPath, 'hist.png')
plt.savefig(rltpath)
plt.show()
def gen_eventdict(sourcePath, saveimg=True):
k, errEvents = 0, []
@ -358,9 +516,7 @@ def gen_eventdict(sourcePath, saveimg=True):
f.write(line + '\n')
def test_one2one():
def init_std_evt_dict():
'''==== 0. 生成事件列表和对应的 Barcodes列表 ==========='''
bcdList, event_spath = [], []
for evtpath in eventSourcePath:
@ -383,10 +539,33 @@ def test_one2one():
print("eventList have generated and saved!")
'''==== 3. 1:1性能评估 ==============='''
rltdata = one2one_simi()
compute_precise_recall(rltdata)
def test_one2one():
'''1:1性能评估'''
# 1. 只需运行一次，生成事件字典和相应的标准特征库字典
# init_std_evt_dict()
# 2. 基于事件barcode集和标准库barcode交集构造事件集合
evtList, evtDict, stdDict = build_std_evt_dict()
rltdata = one2one_simi(evtList, evtDict, stdDict)
one2one_pr(rltdata)
def test_one2SN():
'''1:SN性能评估'''
# 1. 只需运行一次，生成事件字典和相应的标准特征库字典
# init_std_evt_dict()
# 2. 事件barcode集和标准库barcode求交集
evtList, evtDict, stdDict = build_std_evt_dict()
one2SN_pr(evtList, evtDict, stdDict)
if __name__ == '__main__':
'''
@ -402,7 +581,7 @@ if __name__ == '__main__':
stdSamplePath = r"\\192.168.1.28\share\数据\已完成数据\展厅数据\v1.0\比对数据\整理\zhantingBase"
stdBarcodePath = r"D:\exhibition\dataset\bcdpath"
stdFeaturePath = r"D:\exhibition\dataset\feats"
stdFeaturePath = r"\\192.168.1.28\share\数据\已完成数据\比对数据\barcode\all_totalBarocde\features_json\v11_barcode_11592"
# eventSourcePath = [r'D:\exhibition\images\20241202']
# eventSourcePath = [r"\\192.168.1.28\share\测试视频数据以及日志\各模块测试记录\展厅测试\1129_展厅模型v801测试组测试"]
@ -419,6 +598,8 @@ if __name__ == '__main__':
os.makedirs(similPath)
test_one2one()
# test_one2SN()