Camera Capture

A tethered product-photography pipeline for an e-commerce shop. Hit ENTER in the terminal: a Canon EOS M50 fires, the background is removed, the product is centered and watermarked, and a 1500×1500 JPEG lands on disk ready to upload.

Why I built this

Every new product in an online shop needs catalog photos: clean white background, centered, consistent crop, watermarked. Doing this by hand in Lightroom for a batch of products eats a chunk of time. I wanted three keystrokes and a coffee instead.

How fast it is

The three photos in assets/readme/buffer_stop_*.jpg were captured in a single session in about 20 seconds. Each cycle (shutter, USB download, background removal, center, watermark, save) takes roughly seven seconds on a MacBook Pro with an M4 chip. I rotate the product on the turntable between shots and press ENTER. That is the whole loop.

Before / after

Raw images are straight off the 24 MP sensor. Processed ones are what the tool writes: 1500×1500, white background, centered, watermarked. Note the third shot was intentionally taken at an angle. The pipeline does not "straighten" it, because a three-quarter view is exactly what you want on a product page.

Raw	Processed

How it works

One pass, full sensor resolution, one final resize. No SaaS calls, nothing leaves the machine.

1. gphoto2 triggers the shutter and pulls a 6000×4000 JPEG over USB.
2. rembg (model u2netp) runs locally on CPU and produces an alpha mask.
3. Small mask blobs (dust, paper edges) are filtered out by area.
4. Shadows visible through gaps in the product are lifted to white, while the gray plastic body itself is left alone.
5. Bounding box, square crop with a fixed margin, resize to 1500×1500.
6. Unsharp mask on the product only, never on the background, so no halo on alpha edges.
7. Semi-transparent TRIXBRIX.eu watermark in the bottom right.
8. Both the raw camera JPEG and the final 1500×1500 are saved to photos/<product-name>/.
9. Optionally, the processed JPEG is pushed straight to a companion Rails app ("Automat") over a token-authenticated HTTP API, so the shop's photo-studio view updates live as you shoot. The raw stays local — only the finished file goes over the wire.

Run it

source .venv/bin/activate
python3 main.py

Interactive controls:

• ENTER take a shot
• n change the session name (also opens a new Automat session)
• u toggle upload to Automat
• q quit

Background removal is always on — there is no toggle. The raw camera JPEG is saved next to the processed one, so if rembg ever botches a shot you still have the untouched original.

Or re-process an existing file without the camera:

python3 main.py --input some_photo.jpg --name my_product

Pass --no-upload to skip the Automat upload for a one-off run, or unset AUTOMAT_TOKEN to disable it permanently.

Optional: upload to Automat (Rails)

The capture loop can push each processed JPEG to a companion Rails app that hosts the catalog. The product folder name maps to a product record on the Rails side; sessions are deduplicated per product per day, so restarting the script keeps appending to the same gallery.

Copy .env.example to .env and fill in:

AUTOMAT_URL=http://localhost:3000
AUTOMAT_TOKEN=<bearer token from Rails credentials>
AUTOMAT_UPLOAD_ENABLED=true

When the token is set, upload is on by default. The shutter announces the filename to Automat first (so the UI shows a spinner placeholder immediately), then the local pipeline runs, then the processed JPEG is attached to that record.

Stack

Python 3.12 + python-gphoto2, rembg, onnxruntime, Pillow, numpy, scipy.ndimage, rich, plus requests and python-dotenv for the optional Automat upload. No GPU, no cloud inference, runs offline (the upload is the only network hop, and it's local-LAN by default).