Fix distorted people in iPhone videos with rotation metadata

[hugleecool]

Issue
When processing iPhone videos with rotation metadata (especially rotation=-90), people appear horizontally compressed in the output frames. This affects primarily videos recorded with iPhone's front camera in portrait mode.
Root Cause
The current implementation ignores rotation metadata in video files. iPhone videos are stored in landscape format (e.g., 1280×720) with rotation=-90 metadata. While media players automatically apply this rotation, the current code doesn't account for it.
To verify, check any iPhone portrait video metadata:

Solution
Modified video2frames function in preprocess_video.py to:
Detect rotation metadata using ffprobe
For rotation=-90 videos, let ffmpeg handle rotation automatically
Calculate output dimensions correctly based on rotation

Key Changes

# Check for rotation metadata
cmd = f"ffprobe -v error -select_streams v:0 -show_entries stream=side_data_list -of json {video_path}"
result = subprocess.run(cmd, shell=True, capture_output=True, text=True)
info = json.loads(result.stdout)

# For rotation=-90 (iPhone portrait videos)
if rotation == -90:
    # Let ffmpeg handle rotation automatically
    w = H // n_downsample  # Use H as width since ffmpeg will auto-rotate
    h = W // n_downsample  # Use W as height

This ensures proper aspect ratio is maintained for all videos with rotation metadata, especially from mobile devices.

[ShenhanQian]

Thanks for reporting this bug. Can you please share where the variable rotation is defined?

[hugleecool]

I use ffprobe

    # 使用子进程直接调用ffprobe来获取旋转信息
    cmd = f"ffprobe -v error -select_streams v:0 -show_entries stream=side_data_list -of json {video_path}"
    result = subprocess.run(cmd, shell=True, capture_output=True, text=True)
    info = json.loads(result.stdout)
    
    # 检查是否有旋转元数据
    if 'streams' in info and len(info['streams']) > 0:
        stream = info['streams'][0]
        if 'side_data_list' in stream:
            for side_data in stream['side_data_list']:
                if 'rotation' in side_data:
                    rotation = float(side_data['rotation'])
                    print(f'检测到视频旋转元数据: {rotation}度')

[ShenhanQian]

Thanks for the more detailed version. Would you be up for opening a pull request for this fix?

[hugleecool]

ok, will try, first time use git