Tiktok-Talent-Info/utils/video_processing.py

import cv2
import os
import subprocess
import numpy as np
from PIL import Image
from pydub import AudioSegment
from decord import VideoReader, cpu
from concurrent.futures import ThreadPoolExecutor, as_completed
from multiprocessing import Pool, cpu_count

# def split_video_into_segments(video_path, segment_duration=30):
#     """
#     Splits a video into segments of a specified duration using FFmpeg.
#     """
#     output_dir = "/tmp/video_segments"
#     os.makedirs(output_dir, exist_ok=True)

#     # Calculate total duration of the video
#     cap = cv2.VideoCapture(video_path)
#     fps = int(cap.get(cv2.CAP_PROP_FPS))
#     total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
#     total_duration = total_frames / fps
#     cap.release()

#     segments = []
#     for start_time in range(0, int(total_duration), segment_duration):
#         segment_file = os.path.join(output_dir, f"segment_{start_time}.mp4")
#         command = [
#             "ffmpeg", "-y", 
#             "-i", video_path,
#             "-ss", str(start_time),
#             "-t", str(segment_duration),
#             "-c", "copy", segment_file
#         ]
#         subprocess.run(command, check=True)
#         segments.append(segment_file)

#     print(f"segments: \n", segments)
#     return segments

# def split_video_into_segments(video_path, segment_duration=30):    # slow
#     """
#     Splits a video into segments of a specified duration using FFmpeg's segment muxer.
#     """
#     output_dir = "/tmp/video_segments"
#     os.makedirs(output_dir, exist_ok=True)

#     segment_file_pattern = os.path.join(output_dir, "segment_%03d.mp4")

#     command = [
#         "ffmpeg", "-y",
#         "-i", video_path,
#         "-f", "segment",
#         "-segment_time", str(segment_duration),
#         "-c", "copy",
#         "-reset_timestamps", "1",
#         segment_file_pattern
#     ]
#     subprocess.run(command, check=True)

#     segments = sorted([os.path.join(output_dir, f) for f in os.listdir(output_dir) if f.startswith("segment_")])
#     print(f"segments: \n", segments)
#     return segments

def split_video_into_segments(video_path, segment_duration=30):
    """
    Splits a video into segments of a specified duration using FFmpeg.
    """
    output_dir = "/tmp/video_segments"
    os.makedirs(output_dir, exist_ok=True)

    # Calculate total duration of the video
    cap = cv2.VideoCapture(video_path)
    fps = int(cap.get(cv2.CAP_PROP_FPS))
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    total_duration = total_frames / fps
    cap.release()

    segments = []

    def extract_segment(start_time):
        segment_file = os.path.join(output_dir, f"segment_{start_time}.mp4")
        command = [
            "ffmpeg", "-y", 
            "-i", video_path,
            "-ss", str(start_time),
            "-t", str(segment_duration),
            "-c", "copy", segment_file
        ]
        subprocess.run(command, check=True)
        return segment_file

    with ThreadPoolExecutor() as executor:
        futures = [
            executor.submit(extract_segment, start_time)
            for start_time in range(0, int(total_duration), segment_duration)
        ]
        for future in as_completed(futures):
            segments.append(future.result())

    print(f"segments: \n", segments)
    return segments

def extract_audio_from_video(video_path):
    """
    Extract audio from video using pydub and save as a temporary audio file.
    """
    print("Audio extraction started...")
    audio = AudioSegment.from_file(video_path)
    print("Audio extraction completed.")
    audio_path = "/tmp/temp_audio.wav"
    audio.export(audio_path, format="wav")
    print(f"Audio extracted and saved to: {audio_path}")
    return audio_path

############################################################################################################
# optical motion, multithread, calculates motion between consecutive frames using dense optical flow (Farneback) only
# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=1):
#     """
#     Extracts key frames from a video based on motion intensity.
#     """
#     def calculate_motion(frame_pair):
#         """
#         Calculates motion between two consecutive frames using optical flow.
#         """
#         prev_gray, current_frame = frame_pair
#         current_gray = cv2.cvtColor(current_frame, cv2.COLOR_BGR2GRAY)
#         flow = cv2.calcOpticalFlowFarneback(prev_gray, current_gray, None, 0.5, 3, 15, 3, 5, 1.2, 0)
#         motion = np.sum(flow ** 2)
#         return motion, current_gray

#     # Load video frames using Decord
#     video = VideoReader(video_path, ctx=cpu(0))
#     frames_batch = video.get_batch(range(0, len(video), frame_interval)).asnumpy()

#     # Resize frames for faster processing
#     frames = [cv2.resize(frame, (frame.shape[1] // 2, frame.shape[0] // 2)) for frame in frames_batch]

#     # Initialize the first frame
#     prev_gray = cv2.cvtColor(frames[0], cv2.COLOR_BGR2GRAY)
#     frame_pairs = [(prev_gray, frames[i]) for i in range(1, len(frames))]

#     # Calculate motion statistics
#     motion_values = []
#     with ThreadPoolExecutor() as executor:
#         motion_results = list(executor.map(calculate_motion, frame_pairs))
#     motion_values = [motion for motion, _ in motion_results]

#     # Calculate threshold statistically
#     motion_mean = np.mean(motion_values)
#     motion_std = np.std(motion_values)
#     threshold = motion_mean + sigma_multiplier * motion_std

#     # Extract key frames based on motion threshold
#     key_frames = []
#     for i, (motion, frame) in enumerate(zip(motion_values, frames[1:])):
#         if motion > threshold and len(key_frames) < max_frames:
#             img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
#             key_frames.append(img)

#     return key_frames

############################################################################################################
# multithreading with bactch
# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=1):
#     """
#     Extracts key frames from a video based on motion intensity.
#     Optimized for speed and efficiency.
#     """
#     def calculate_motion(frame_pair):
#         """
#         Calculates motion between two consecutive frames using optical flow.
#         """
#         prev_gray, current_frame = frame_pair
#         current_gray = cv2.cvtColor(current_frame, cv2.COLOR_BGR2GRAY)
#         flow = cv2.calcOpticalFlowFarneback(prev_gray, current_gray, None, 0.5, 3, 15, 3, 5, 1.2, 0)
#         motion = np.sum(flow ** 2)
#         return motion, current_gray

#     # Load video frames using Decord with reduced resolution
#     video = VideoReader(video_path, ctx=cpu(0))
#     total_frames = len(video)
#     frame_indices = range(0, total_frames, frame_interval)

#     # Process frames in smaller batches to reduce memory usage
#     batch_size = 100  
#     motion_values = []

#     for batch_start in range(0, len(frame_indices), batch_size):
#         batch_end = min(batch_start + batch_size, len(frame_indices))
#         batch_indices = frame_indices[batch_start:batch_end]
#         frames_batch = video.get_batch(batch_indices).asnumpy()

#         # Resize frames for faster processing
#         frames = [cv2.resize(frame, (frame.shape[1] // 2, frame.shape[0] // 2)) for frame in frames_batch]

#         # Initialize the first frame in the batch
#         prev_gray = cv2.cvtColor(frames[0], cv2.COLOR_BGR2GRAY)
#         frame_pairs = [(prev_gray, frames[i]) for i in range(1, len(frames))]

#         # Calculate motion statistics for the batch
#         with ThreadPoolExecutor() as executor:
#             motion_results = list(executor.map(calculate_motion, frame_pairs))
#         batch_motion_values = [motion for motion, _ in motion_results]
#         motion_values.extend(batch_motion_values)

#         # Update the previous frame for the next batch
#         prev_gray = cv2.cvtColor(frames[-1], cv2.COLOR_BGR2GRAY)

#     # Calculate threshold statistically
#     motion_mean = np.mean(motion_values)
#     motion_std = np.std(motion_values)
#     threshold = motion_mean + sigma_multiplier * motion_std

#     # Extract key frames based on motion threshold
#     key_frames = []
#     for i, (motion, frame) in enumerate(zip(motion_values, frames[1:])):
#         if motion > threshold and len(key_frames) < max_frames:
#             img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
#             key_frames.append(img)

#     return key_frames

############################################################################################################
# multiprocessing
# def calculate_motion(frame_pair):
#     """
#     Calculates motion between two consecutive frames using optical flow.
#     """
#     prev_gray, current_gray = frame_pair
#     flow = cv2.calcOpticalFlowFarneback(prev_gray, current_gray, None, 0.5, 3, 15, 3, 5, 1.2, 0)
#     motion = np.sum(flow ** 2)
#     return motion

# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=1):
#     """
#     Extracts key frames from a video based on motion intensity.
#     Optimized for speed and efficiency.
#     """
#     # Load video frames using Decord with reduced resolution
#     video = VideoReader(video_path, ctx=cpu(0))
#     total_frames = len(video)
#     frame_indices = range(0, total_frames, frame_interval)

#     # Read all frames and resize them for faster processing
#     frames = video.get_batch(frame_indices).asnumpy()
#     frames = [cv2.resize(frame, (frame.shape[1] // 2, frame.shape[0] // 2)) for frame in frames]

#     # Convert all frames to grayscale in one go
#     grayscale_frames = [cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) for frame in frames]

#     # Calculate motion between consecutive frames using multiprocessing
#     frame_pairs = list(zip(grayscale_frames[:-1], grayscale_frames[1:]))
#     with Pool(cpu_count()) as pool:
#         motion_values = pool.map(calculate_motion, frame_pairs)

#     # Calculate threshold statistically
#     motion_mean = np.mean(motion_values)
#     motion_std = np.std(motion_values)
#     threshold = motion_mean + sigma_multiplier * motion_std

#     # Extract key frames based on motion threshold
#     key_frames = []
#     for i, motion in enumerate(motion_values):
#         if motion > threshold and len(key_frames) < max_frames:
#             img = Image.fromarray(cv2.cvtColor(frames[i + 1], cv2.COLOR_BGR2RGB))
#             key_frames.append(img)

#     return key_frames

############################################################################################################
# faster optical flow, more aggressive downscaling and frame skipping, looking for motion peaks, uses both dense optical flow and includes additional peak detection logic
# def calculate_motion(frames):
#     """
#     Calculate motion metrics using frame differencing and sparse optical flow
#     Returns a list of motion intensity values
#     """
#     if len(frames) < 2:
#         return []
    
#     # Convert all frames to grayscale at once
#     gray_frames = [cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) for frame in frames]
    
#     # Parameters for ShiTomasi corner detection and optical flow
#     feature_params = dict(maxCorners=100, qualityLevel=0.3, minDistance=7, blockSize=7)
#     lk_params = dict(winSize=(15,15), maxLevel=2, 
#                     criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 0.03))
    
#     motion_metrics = []
#     prev_frame = gray_frames[0]
#     prev_pts = cv2.goodFeaturesToTrack(prev_frame, mask=None, **feature_params)
    
#     for i in range(1, len(gray_frames)):
#         curr_frame = gray_frames[i]
        
#         # Calculate dense optical flow (Farneback)
#         flow = cv2.calcOpticalFlowFarneback(prev_frame, curr_frame, None, 0.5, 3, 15, 3, 5, 1.2, 0)
#         magnitude = np.sqrt(flow[...,0]**2 + flow[...,1]**2)
#         motion_metrics.append(np.mean(magnitude))
        
#         prev_frame = curr_frame
    
#     return motion_metrics

# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=5):
#     # Load video with reduced resolution
#     video = VideoReader(video_path, ctx=cpu(0))
#     total_frames = len(video)
#     frame_indices = range(0, total_frames, frame_interval)
    
#     # Read and resize all frames at once
#     frames = video.get_batch(frame_indices).asnumpy()
#     frames = np.array([cv2.resize(frame, (frame.shape[1]//4, frame.shape[0]//4)) for frame in frames])
    
#     # Calculate motion metrics
#     motion_values = calculate_motion(frames)
    
#     if not motion_values:
#         return []
    
#     # Adaptive thresholding
#     mean_motion = np.mean(motion_values)
#     std_motion = np.std(motion_values)
#     threshold = mean_motion + sigma_multiplier * std_motion
    
#     # Find peaks in motion values
#     key_frame_indices = []
#     for i in range(1, len(motion_values)-1):
#         if motion_values[i] > threshold and \
#            motion_values[i] > motion_values[i-1] and \
#            motion_values[i] > motion_values[i+1]:
#             key_frame_indices.append(i+1)  # +1 because motion is between frames
    
#     # Select top frames by motion intensity
#     if len(key_frame_indices) > max_frames:
#         sorted_indices = sorted(key_frame_indices, key=lambda x: motion_values[x-1], reverse=True)
#         key_frame_indices = sorted_indices[:max_frames]
#         key_frame_indices.sort()
    
#     # Convert to PIL Images
#     key_frames = [Image.fromarray(cv2.cvtColor(frames[i], cv2.COLOR_BGR2RGB)) 
#                  for i in key_frame_indices]
    
#     return key_frames


############################################################################################################
# RAFT Optical Flow
# import torch
# import torchvision.models.optical_flow as of
# from torch.nn.parallel import DataParallel

# def pad_to_multiple_of_8(frame):
#     """
#     Pads the frame dimensions to the nearest multiple of 8.
#     """
#     h, w, _ = frame.shape
#     pad_h = (8 - h % 8) % 8
#     pad_w = (8 - w % 8) % 8
#     return cv2.copyMakeBorder(frame, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=[0, 0, 0])

# def compute_raft_flow_batch(frame_batch, raft_model):
#     """
#     Computes optical flow for a batch of frames using the RAFT model.
#     """
#     # Pad frames to make dimensions divisible by 8
#     frame_batch = [pad_to_multiple_of_8(frame) for frame in frame_batch]

#     # Convert frames to tensors and normalize
#     frame_tensors = torch.stack([torch.from_numpy(frame).permute(2, 0, 1).float().cuda() / 255.0 for frame in frame_batch])

#     # Compute optical flow for the batch
#     with torch.no_grad():
#         flows = raft_model(frame_tensors[:-1], frame_tensors[1:])

#     # Calculate motion magnitude for each flow
#     motions = [np.sum(flow.cpu().numpy() ** 2) for flow in flows]
#     return motions

# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=1, batch_size=128):
#     """
#     Extracts key frames from a video based on motion intensity using RAFT for optical flow.
#     Utilizes multiple GPUs and processes frames in batches.
#     """
#     # Load RAFT model and wrap it with DataParallel for multi-GPU support
#     print("Loading RAFT model...")
#     raft_model = of.raft_large(pretrained=True).cuda()
#     if torch.cuda.device_count() > 1:
#         print(f"Using {torch.cuda.device_count()} GPUs!")
#         raft_model = DataParallel(raft_model)

#     # Load video frames using Decord with reduced resolution
#     video = VideoReader(video_path, ctx=cpu(0))
#     total_frames = len(video)
#     frame_indices = range(0, total_frames, frame_interval)

#     # Read all frames and resize them for faster processing
#     frames = video.get_batch(frame_indices).asnumpy()
#     frames = [cv2.resize(frame, (frame.shape[1] // 2, frame.shape[0] // 2)) for frame in frames]

#     # Calculate motion between consecutive frames using RAFT in batches
#     motion_values = []
#     print(f"The total number of frames: {len(frames)}")
#     for batch_start in range(1, len(frames), batch_size):
#         batch_end = min(batch_start + batch_size, len(frames))
#         batch_frames = frames[batch_start - 1:batch_end]
#         batch_motions = compute_raft_flow_batch(batch_frames, raft_model)
#         motion_values.extend(batch_motions)

#     # Calculate threshold statistically
#     motion_mean = np.mean(motion_values)
#     motion_std = np.std(motion_values)
#     threshold = motion_mean + sigma_multiplier * motion_std

#     # Extract key frames based on motion threshold
#     key_frames = []
#     for i, motion in enumerate(motion_values):
#         if motion > threshold and len(key_frames) < max_frames:
#             img = Image.fromarray(cv2.cvtColor(frames[i + 1], cv2.COLOR_BGR2RGB))
#             key_frames.append(img)

#     return key_frames

############################################################################################################
# Histogram Difference
# def calculate_histogram_difference(frame_pair):
#     """
#     Calculates the difference between two consecutive frames using color histograms.
#     """
#     frame1, frame2 = frame_pair

#     # Calculate histograms for each frame
#     hist1 = cv2.calcHist([frame1], [0, 1, 2], None, [8, 8, 8], [0, 256, 0, 256, 0, 256])
#     hist2 = cv2.calcHist([frame2], [0, 1, 2], None, [8, 8, 8], [0, 256, 0, 256, 0, 256])

#     # Normalize histograms
#     cv2.normalize(hist1, hist1)
#     cv2.normalize(hist2, hist2)

#     # Calculate histogram difference using Chi-Squared distan
#     difference = cv2.compareHist(hist1, hist2, cv2.HISTCMP_CHISQR)
#     return difference

# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=1):
#     """
#     Extracts key frames from a video based on histogram differences.
#     Optimized for speed and efficiency.
#     """
#     # Load video frames using Decord with reduced resolution
#     video = VideoReader(video_path, ctx=cpu(0))
#     total_frames = len(video)
#     frame_indices = range(0, total_frames, frame_interval)

#     # Read all frames and resize them for faster processing
#     frames = video.get_batch(frame_indices).asnumpy()
#     frames = [cv2.resize(frame, (frame.shape[1] // 2, frame.shape[0] // 2)) for frame in frames]

#     # Calculate histogram differences between consecutive frames using multiprocessing
#     frame_pairs = list(zip(frames[:-1], frames[1:]))
#     with Pool(cpu_count()) as pool:
#         histogram_differences = pool.map(calculate_histogram_difference, frame_pairs)

#     # Calculate threshold statistically
#     diff_mean = np.mean(histogram_differences)
#     diff_std = np.std(histogram_differences)
#     threshold = diff_mean + sigma_multiplier * diff_std

#     # Extract key frames based on histogram difference threshold
#     key_frames = []
#     for i, difference in enumerate(histogram_differences):
#         if difference > threshold and len(key_frames) < max_frames:
#             img = Image.fromarray(cv2.cvtColor(frames[i + 1], cv2.COLOR_BGR2RGB))
#             key_frames.append(img)

#     return key_frames


############################################################################################################
# faster histogram
# def calculate_histogram_difference(frame1, frame2):
#     """
#     Calculates the difference between two consecutive frames using grayscale histograms.
#     """
#     # Convert frames to grayscale
#     gray1 = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
#     gray2 = cv2.cvtColor(frame2, cv2.COLOR_BGR2GRAY)

#     # Calculate histograms with fewer bins (e.g., 16 bins)
#     hist1 = cv2.calcHist([gray1], [0], None, [16], [0, 256])
#     hist2 = cv2.calcHist([gray2], [0], None, [16], [0, 256])

#     # Normalize histograms
#     cv2.normalize(hist1, hist1)
#     cv2.normalize(hist2, hist2)

#     # Calculate histogram difference using Chi-Squared distance
#     difference = cv2.compareHist(hist1, hist2, cv2.HISTCMP_CHISQR)
#     return difference

# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=10):
#     """
#     Extracts key frames from a video based on histogram differences.
#     Optimized for speed by reducing histogram complexity and skipping frames.
#     """
#     # Load video frames using Decord with reduced resolution
#     video = VideoReader(video_path, ctx=cpu(0))
#     total_frames = len(video)
#     frame_indices = range(0, total_frames, frame_interval)

#     # Read all frames and resize them for faster processing
#     frames = video.get_batch(frame_indices).asnumpy()
#     frames = [cv2.resize(frame, (frame.shape[1] // 2, frame.shape[0] // 2)) for frame in frames]

#     # Calculate histogram differences between consecutive frames
#     histogram_differences = []
#     for i in range(1, len(frames)):
#         difference = calculate_histogram_difference(frames[i - 1], frames[i])
#         histogram_differences.append(difference)

#     # Calculate threshold statistically
#     diff_mean = np.mean(histogram_differences)
#     diff_std = np.std(histogram_differences)
#     threshold = diff_mean + sigma_multiplier * diff_std

#     # Extract key frames based on histogram difference threshold
#     key_frames = []
#     for i, difference in enumerate(histogram_differences):
#         if difference > threshold and len(key_frames) < max_frames:
#             img = Image.fromarray(cv2.cvtColor(frames[i + 1], cv2.COLOR_BGR2RGB))
#             key_frames.append(img)

#     return key_frames

############################################################################################################
# faster histogram with batch
def calculate_histogram_difference_batch(frame_batch):
    """
    Calculates histogram differences for a batch of frames.
    """
    # Convert frames to grayscale
    gray_frames = [cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) for frame in frame_batch]

    # Calculate histograms for all frames in the batch
    histograms = [cv2.calcHist([gray], [0], None, [16], [0, 256]) for gray in gray_frames]
    for hist in histograms:
        cv2.normalize(hist, hist)

    # Calculate histogram differences between consecutive frames
    differences = []
    for i in range(1, len(histograms)):
        difference = cv2.compareHist(histograms[i - 1], histograms[i], cv2.HISTCMP_CHISQR)
        differences.append(difference)

    return differences

def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=10, batch_size=16):
    """
    Extracts key frames from a video based on histogram differences.
    Uses batch processing for faster computation.
    """
    # Load video frames using Decord with reduced resolution
    video = VideoReader(video_path, ctx=cpu(0))
    total_frames = len(video)
    print(f"All total frames: {total_frames}")
    frame_indices = range(0, total_frames, frame_interval)

    # Read all frames and resize them for faster processing
    frames = video.get_batch(frame_indices).asnumpy()
    frames = [cv2.resize(frame, (frame.shape[1] // 2, frame.shape[0] // 2)) for frame in frames]

    # Process frames in batches
    histogram_differences = []
    print(f"The total number of frames: {len(frames)}")
    for batch_start in range(0, len(frames), batch_size):
        batch_end = min(batch_start + batch_size + 1, len(frames))  # +1 to include the next frame for difference
        batch_frames = frames[batch_start:batch_end]
        batch_differences = calculate_histogram_difference_batch(batch_frames)
        histogram_differences.extend(batch_differences)

    # Calculate threshold statistically
    diff_mean = np.mean(histogram_differences)
    diff_std = np.std(histogram_differences)
    threshold = diff_mean + sigma_multiplier * diff_std

    # Extract key frames based on histogram difference threshold
    key_frames = []
    for i, difference in enumerate(histogram_differences):
        if difference > threshold and len(key_frames) < max_frames:
            img = Image.fromarray(cv2.cvtColor(frames[i + 1], cv2.COLOR_BGR2RGB))
            key_frames.append(img)

    return key_frames


############################################################################################################
# faster faster histogram
# def calculate_frame_difference(frame1, frame2):
#     """
#     Ultra-fast frame difference calculation using downscaled grayscale and absolute pixel differences.
#     """
#     # Convert to grayscale and downscale further
#     gray1 = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
#     gray2 = cv2.cvtColor(frame2, cv2.COLOR_BGR2GRAY)
    
#     # Downscale to tiny images (e.g., 16x16) for fast comparison
#     tiny1 = cv2.resize(gray1, (16, 16))
#     tiny2 = cv2.resize(gray2, (16, 16))
    
#     # Calculate normalized absolute difference
#     diff = cv2.absdiff(tiny1, tiny2)
#     return np.mean(diff) / 255.0  # Normalize to [0,1]

# def save_key_frames(key_frames, output_dir="key_frames", prefix="frame"):
#     """
#     Saves key frames to disk as JPEG images.
#     """
#     if not os.path.exists(output_dir):
#         os.makedirs(output_dir)
    
#     saved_paths = []
#     for i, frame in enumerate(key_frames):
#         frame_path = os.path.join(output_dir, f"{prefix}_{i:04d}.jpg")
#         frame.save(frame_path, quality=85)  # Good quality with reasonable compression
#         saved_paths.append(frame_path)
    
#     return saved_paths

# def extract_motion_key_frames(video_path, max_frames=20, sigma_multiplier=2, frame_interval=15):
#     # Load video with decord (faster than OpenCV)
#     video = VideoReader(video_path, ctx=cpu(0))
#     total_frames = len(video)
    
#     # Pre-calculate frame indices to process
#     frame_indices = range(0, total_frames, frame_interval)
#     frames = video.get_batch(frame_indices).asnumpy()
    
#     # Downscale all frames upfront (much faster than per-frame)
#     frames = [cv2.resize(frame, (frame.shape[1]//4, frame.shape[0]//4)) for frame in frames]
    
#     # Calculate differences (vectorized approach)
#     differences = []
#     prev_frame = frames[0]
#     for frame in frames[1:]:
#         diff = calculate_frame_difference(prev_frame, frame)
#         differences.append(diff)
#         prev_frame = frame
    
#     # Adaptive thresholding
#     diff_mean = np.mean(differences)
#     diff_std = np.std(differences)
#     threshold = diff_mean + sigma_multiplier * diff_std
    
#     # Extract key frames
#     key_frames = []
#     for i, diff in enumerate(differences):
#         if diff > threshold and len(key_frames) < max_frames:
#             img = Image.fromarray(cv2.cvtColor(frames[i+1], cv2.COLOR_BGR2RGB))
#             key_frames.append(img)

#     saved_paths = save_key_frames(key_frames, '../video')
    
#     return key_frames