优化人物淡入淡出时，人脸判断依然有效的情况：
引入sample_rate参数来控制帧率采样，减少不必要的处理。
使用min_confidence参数来过滤低置信度的人脸检测结果。
使用滑动窗口计算置信度的移动平均，以检测淡出效果。fade_threshold参数用于决定何时认为人脸真正淡出。
我们还增加了timeout参数，用于在没有人脸被检测到一定时间后停止搜索。
注意，face_recognition库的face_encodings方法返回的编码并不直接对应置信度，但我们可以利用编码的范数作为置信度的代理。这个范数通常在人脸存在时会相对较小，不存在时较大，因此可以作为置信度的反向指标。
这个版本的代码应该能够更准确地处理淡入淡出效果，并在保持较高检测灵敏度的同时，加快处理速度。不过，你可能需要根据实际情况调整sample_rate、min_confidence、fade_threshold和timeout等参数。
完整代码：
# python data_utils/pre_video/cut_crop_fps.py
import cv2
import math
import numpy as np
import face_recognition
from moviepy.editor import VideoFileClip, concatenate_videoclips
from tqdm import tqdm
def find_host_face_location(video_path):
""" 在视频的前几秒内检测并返回主持人面部的大致位置 """
cap = cv2.VideoCapture(video_path)
found_face = False
host_face_location = None
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
# 缩小帧尺寸以加快处理速度
small_frame = cv2.resize(frame, (0, 0), fx=0.25, fy=0.25)
rgb_small_frame = small_frame[:, :, ::-1]
# 检测人脸
face_locations = face_recognition.face_locations(rgb_small_frame)
if face_locations:
# 取第一张脸的位置，假设主持人位于视频画面的中心位置附近
host_face_location = face_locations[0]
# 将位置放大回原始大小
host_face_location = (host_face_location[0]*4, host_face_location[1]*4, host_face_location[2]*4, host_face_location[3]*4)
found_face = True
break
cap.release()
return host_face_location if found_face else None
def calculate_cropping_box(face_location, frame_shape):
""" 根据主持人面部位置计算裁剪框 """
top, right, bottom, left = face_location
center_x, center_y = (left + right) // 2, (top + bottom) // 2
half_width, half_height = 256, 256
left_cropped = max(center_x - half_width, 0)
top_cropped = max(center_y - half_height, 0)
right_cropped = min(center_x + half_width, frame_shape[1])
bottom_cropped = min(center_y + half_height, frame_shape[0])
return (top_cropped, right_cropped, bottom_cropped, left_cropped)
def find_first_last_face(video_path, sample_rate=3, min_confidence=0.5, fade_threshold=0.5, timeout=2.0):
""" 找到视频中第一次和最后一次出现人脸的时间戳，优化处理速度和淡出效果 """
cap = cv2.VideoCapture(video_path)
fps = cap.get(cv2.CAP_PROP_FPS)
first_face_time = None
last_face_time = 0
last_face_detected_time = 0
timeout_counter = 0.0
face_confidences = []
timestamps = []
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
# 按照sample_rate进行帧率采样
if len(timestamps) > 0 and (len(timestamps) % sample_rate) != 0:
continue
timestamp = cap.get(cv2.CAP_PROP_POS_MSEC) / 1000 # Convert to seconds
# 缩小帧尺寸以加快处理速度
small_frame = cv2.resize(frame, (0, 0), fx=0.25, fy=0.25)
rgb_small_frame = small_frame[:, :, ::-1]
# 检测人脸及其置信度
face_locations = face_recognition.face_locations(rgb_small_frame, model='hog') # 可以选择 'cnn' 或 'hog'
face_encodings = face_recognition.face_encodings(rgb_small_frame, face_locations)
face_confidence_scores = [np.linalg.norm(face_encoding) for face_encoding in face_encodings]
if face_locations:
if not first_face_time:
first_face_time = timestamp
last_face_detected_time = timestamp
timeout_counter = 0.0
# 计算平均置信度
avg_confidence = sum(face_confidence_scores) / len(face_confidence_scores)
face_confidences.append(avg_confidence)
timestamps.append(timestamp)
# 检查是否低于置信度阈值
if avg_confidence < min_confidence:
face_confidences[-1] = 0
else:
timeout_counter += 1/fps
face_confidences.append(0)
timestamps.append(timestamp)
if timeout_counter > timeout:
last_face_time = last_face_detected_time
break
cap.release()
# 后处理逻辑：滑动窗口检测强度
window_size = int(timeout * fps / sample_rate)
if window_size > 1:
moving_averages = np.convolve(face_confidences, np.ones(window_size)/window_size, mode='valid')
moving_timestamps = timestamps[window_size-1:]
for i, avg in enumerate(moving_averages):
if avg < fade_threshold and last_face_time == 0:
last_face_time = moving_timestamps[i]
break
# 如果整个视频都没有检测到人脸，设置last_face_time为None
if last_face_time == 0:
last_face_time = None
return first_face_time, last_face_time
def process_video(input_path, output_path):
""" 处理视频，裁剪并调整帧率 """
# 检测主持人面部位置
host_face_location = find_host_face_location(input_path)
if host_face_location is None:
print(f"No face detected in video {input_path}")
return
# 读取视频，获取视频的宽度和高度
clip = VideoFileClip(input_path)
frame_shape = clip.size[::-1] # 电影剪辑的尺寸是(width, height)，我们需要(height, width)
# 计算裁剪框
cropping_box = calculate_cropping_box(host_face_location, frame_shape)
# 找到第一次和最后一次出现人脸的时间
first_face_time, last_face_time = find_first_last_face(input_path)
print(f"First face time: {first_face_time}, Last face time: {last_face_time}")
# 裁剪视频以保留第一次和最后一次出现人脸的部分
start_trim = math.ceil(first_face_time) # 向上取整
end_trim = math.floor(last_face_time) # 向下取整
print(f"Start trim: {start_trim}, End trim: {end_trim}")
trimmed_clip = clip.subclip(start_trim, end_trim)
# 裁剪视频
cropped_clip = trimmed_clip.crop(x1=cropping_box[3], y1=cropping_box[0], x2=cropping_box[1], y2=cropping_box[2])
cropped_clip = cropped_clip.resize((512, 512))
# 调整帧率
cropped_clip = cropped_clip.set_fps(25)
# 保存最终视频
cropped_clip.write_videofile(output_path, codec='libx264', audio_codec='aac')
# 清理资源
cropped_clip.close()
if __name__ == "__main__":
for i in tqdm(range(1, 76), desc="Processing videos"):
print("处理第", i, "个视频")
input_path = f"data/dataset/{i}/{i}.mp4"
output_path = f"data/dataset/{i}/{i}_fcc.mp4"
process_video(input_path, output_path)