What it does

Autocrop takes a folder of photos, finds the largest face in each, and outputs a centered square crop. It handles the entire pipeline: input scanning, face detection via OpenCV’s Haar cascades, resizing, and sorting failures into a reject folder. There’s both a Python API (returns NumPy arrays) and a CLI with sensible defaults.

The interesting bit

The author is admirably honest: this is “essentially a CLI wrapper around a single OpenCV function.” The value isn’t the detection algorithm—it’s the ergonomics. In-place destructive cropping with confirmation prompts, a --no-resize flag for pixel purists, and a documented ffmpeg pipeline for video frame extraction. The boring parts (file I/O, format support for 20+ image types, cross-platform testing) are where the work actually went.

Key highlights

• CLI and Python API both supported; outputs raw NumPy arrays for downstream processing
• 20+ image formats including WebP, TIFF, and ancient formats like PCX and SGI
• Reject folder for images where no face is detected, so failed crops don’t silently disappear
• --no-resize preserves original pixels; --facePercent controls face-to-frame ratio
• Actively maintained with CI across Python 3.7–3.10 on Linux, macOS, and Windows

Caveats

• Conda packaging is stalled due to OpenCV workflow complexity (issue #7)
• Only the first frame of animated GIFs is processed
• Defaults to destructive in-place cropping if you forget the output folder

Verdict

Worth a look if you process ID photos, profile pictures, or video frames in bulk. Skip it if you need fine-grained face alignment, multiple-face handling, or non-square aspect ratios—this centers on the single largest face and outputs squares.