def main():
flg_det = PdxDet(model_path="../model/best/flg_det/", bs=8)
for vid_name in os.listdir(args.input):
print(vid_name)
frame_data = []
names = []
name = vid_name.split(".")[0]
video = Stream(
osp.join(args.input, vid_name),
osp.join(args.ann, name + ".txt"),
itv_sparse=0,
itv_dense=5,
)
for frame_idx, img in tqdm(video):
# print(idx)
# # 第一种方式推理类负责组batch
frames, idxs, bbs = flg_det.add(img, frame_idx)
# print(bbs)
for f, idx, bb in zip(frames, idxs, bbs):
if len(bb) != 0:
bb = bb[0] # 这个网络最多检出一个起落架
save_path = osp.join(args.output,
"{}-{}.png".format(name, idx))
try:
cv2.imwrite(save_path, crop(f, bb.square(256)))
except:
pass
def main():
flg_det = PdxDet(model_path="../model/best/flg_det/", bs=args.bs)
for vid_name in os.listdir(args.input):
print(vid_name)
name = vid_name.split(".")[0]
video = Stream(
osp.join(args.input, vid_name),
# osp.join(args.ann, name + ".txt"),
itv_sparse=2,
# itv_dense=2,
)
os.mkdir(osp.join(args.output, "p" + vid_name))
for frame_idx, img in tqdm(video, miniters=args.bs):
frames, idxs, bbs = flg_det.add(img, frame_idx)
for f, idx, bb in zip(frames, idxs, bbs):
# print(bb)
if len(bb) != 0:
bb = bb[0] # 这个网络最多检出一个起落架
save_path = osp.join(
args.output,
"p" + vid_name,
"{}-{}.png".format(name, idx),
)
try:
cv2.imwrite(save_path, crop(f, bb.square(256)))
except:
pass
def main():
# 1. 定义模型对象
flg_det = PdxDet(model_path="../model/best/flg_det/", bs=4)
person_det = PdxDet(model_path="../model/best/person_det_yolov3", bs=8, autoflush=False)
for vid_name in os.listdir(args.input):
print(vid_name)
os.mkdir(osp.join(args.output, vid_name))
video = Stream(osp.join(args.input, vid_name), osp.join(args.time, os.listdir(args.time)[0]))
# TODO: 研究tqdm需要什么方法显示总数
for fidx, img in video:
# 检测出一个batch的起落架
frames, fidxs, flgs_batch = flg_det.add(img, fidx)
for frame, frame_idx, flgs in zip(frames, fidxs, flgs_batch): # 对这些起落架中的每一个
if len(flgs) != 0:
flg = flgs[0]
person_det.add(crop(frame, flg.square(256)), [flg, frame_idx]) # 添加到检测人的任务list中
if len(flgs_batch) != 0:
print("Gears detected: ", flgs_batch)
if len(person_det.imgs) >= person_det.bs:
count = 0
r = person_det.flush() # 进行人像检测
print("People detected: ", r[2])
for gear_square, info, people in zip(r[0], r[1], r[2]): # 对一个batch中的每一张，每一张可能有多个人
f = frames[count]
flg = info[0]
fid = info[1]
count += 1
for pid, person in enumerate(people):
dbb(f, flg.square(256).transpose(person), "G")
patch = crop(f, flg.square(256).transpose(person).square(128))
if patch.shape == (128, 128, 3):
cv2.imwrite(
osp.join(args.output, vid_name, "{}-{}.png".format(pid, str(fid))), patch
)
# cv2.imshow("img", crop(gear_square, person))
dbb(f, flg, "R")
dpoint(f, flg.center(), "R")
dbb(f, flg.square(256), "B")
#
cv2.imshow("img", f)
cv2.waitKey(2)
import os
import os.path as osp
from tqdm import tqdm
import cv2
from util.util import Stream
vid_path = "/home/2t/plane/视频划分/train"
img_path = "/home/aistudio/plane/temp"
for vid_name in tqdm(os.listdir(vid_path)):
video = Stream(
osp.join(vid_path, vid_name),
itv_dense=12,
)
vid_name = osp.splitext(vid_name)[0]
print(vid_name)
for idx, frame in video:
cv2.imwrite(osp.join(img_path, f"{vid_name}-{str(idx).zfill(5)}.png"), frame)
# input("here")
def main():
# 1. 定义模型对象
flg_det = PdxDet(model_path="../model/best/flg_det/", bs=2)
person_det = PdxDet(model_path="../model/best/person_det_yolov3",
bs=2,
autoflush=False)
person_clas = HumanClas(mode="predict")
for vid_name in os.listdir(args.input):
# print(vid_name)
#os.mkdir(osp.join(args.output, vid_name))
video = Stream(
osp.join(args.input, vid_name),
osp.join(args.time,
os.listdir(args.time)[0]),
itv_sparse=3,
itv_dense=3,
)
mem_len = 8
history = [False for _ in range(mem_len)]
thresh = 0.5
# TODO: 研究tqdm需要什么方法显示总数
#res_f = open(osp.join("H:/W_S/Graduation_Project/plane/time-out", vid_name.rstrip(".mp4") + ".txt"), "w")
for fidx, img in video:
# 检测出一个batch的起落架
frames, fidxs, flgs_batch = flg_det.add(img, fidx)
for frame, frame_idx, flgs in zip(frames, fidxs,
flgs_batch): # 对这些起落架中的每一个
if len(flgs) != 0:
flg = flgs[0]
person_det.add(crop(frame, flg.square(256)),
[flg, frame_idx, frame]) # 添加到检测人的任务list中
# print("Gears detected: ", flgs_batch)
if len(person_det.imgs) >= person_det.bs:
r = person_det.flush() # 进行人像检测
# print("People detected: ", r[2])
for gear_square, info, people in zip(
r[0], r[1], r[2]): # 对一个batch中的每一张，每一张可能有多个人
flg = info[0]
fid = info[1]
f = info[2]
# TODO: 一个batch推理
has_positive = False
for pid, person in enumerate(people):
patch = crop(
f,
flg.square(256).transpose(person).square(64))
res = person_clas.predict(patch)
if res:
has_positive = True
# dbb(f, flg.square(256).transpose(person).region([1.8, 1.8]), "G" if res else "R")
dpoint(f,
flg.square(256).transpose(person).center(),
"G" if res else "R")
for idx in range(mem_len - 1, 0, -1):
history[idx] = history[idx - 1]
history[0] = has_positive
prediction = "Positive" if np.sum(
history) > mem_len * thresh else "Negative"
print(has_positive)
print(history, np.sum(history), prediction)
#print(fid, has_positive, prediction, np.sum(history), file=res_f)
#res_f.flush()
dbb(f, flg, "B")
dpoint(f, flg.center(), "B")
dbb(f, flg.square(256), "B")
cv2.imshow("img", f)
cv2.waitKey()
# cv2.imwrite(
# osp.join(args.output, vid_name, str(fid).zfill(6) + ".png"),
# crop(f, flg.square(300)),
# )
#res_f.close()
shutil.move(osp.join(args.output, vid_name),
osp.join(args.output, "finish"))
###########################################################
# Stream Test
###########################################################
from util.util import Stream
import cv2
vid = Stream(
"/home/aistudio/plane/vid-split/train/15251-上轮挡.mp4",
"/home/aistudio/plane/时间标注/train/15251-上轮挡.txt",
itv_sparse=0,
itv_dense=5,
)
for idx, frame in vid:
cv2.imshow("img", frame)
cv2.waitKey()
import os
import os.path as osp
from tqdm import tqdm
import cv2
from util.util import xml2bb, Stream, crop
vid_dir = "/home/aistudio/plane/视频分类/1920/n/"
bb_dir = "/home/aistudio/plane/视频分类/起落架bb/"
out_dir = "/home/aistudio/plane/视频分类/256-frame/n/"
for name in tqdm(os.listdir(vid_dir)):
name = name.split(".")[0]
bb_path = osp.join(bb_dir, name + ".xml")
if os.path.exists(bb_path):
print(name)
bb = xml2bb(bb_path, "gear")[0]
vid = cv2.VideoCapture(osp.join(vid_dir, name + ".mp4"))
stream = Stream(osp.join(vid_dir, name + ".mp4"), itv_sparse=2)
os.mkdir(osp.join(out_dir, name))
for idx, img in stream:
out_path = osp.join(out_dir, name,
"{}-{}.png".format(name,
str(idx).zfill(6)))
img = crop(img, bb.square(256), do_pad=True)
cv2.imwrite(out_path, img)
import numpy as np
import cv2
from util.util import Stream
video = Stream("/home/2t/plane/视频划分/long/Q216022021.3.20", itv_sparse=1500)
print(video.fps)
prev_diff = 0
print(video.shape)
prev_frame = np.zeros(video.shape)
f = open("stamp.txt", "w")
for idx, curr_frame in video:
curr_diff = np.abs(curr_frame - prev_frame).sum()
print(idx, prev_diff, curr_diff)
if curr_diff > 1000 * prev_diff:
print(int(idx / 25), file=f, end=" ")
print("+++++++++++++++ Starting new video")
f.flush()
prev_diff = curr_diff
cv2.imshow("img", curr_frame)
cv2.waitKey(2)