Augmentation Playground

Interactive browser UI for building Albumentations pipelines — tweak sliders, see results live, copy the exact Python code.

Upload an image, drag a slider, hit apply. The generated A.Compose([...]) snippet in the right panel always matches what you see.

---

Why this exists

Building a good augmentation pipeline usually means writing code, running training, checking if it helped, adjusting, repeating. This tool collapses that loop — you see the effect of every parameter before you commit it to your training script.

It also supports full polygon labels (YOLO segmentation format) — upload a .txt label file alongside your image and the polygon contours transform correctly with every augmentation, so you can verify your geometric transforms aren't breaking your annotations.

---

Features

Side-by-side compare	Original and augmented shown at the same size, same position
Grid ×9	See 9 different random variants at once — great for judging augmentation strength
Polygon label support	Upload YOLO polygon .txt labels, contours drawn and correctly transformed
Toggle overlay	Switch polygon visualization on/off without re-running augmentation
Code export	Generated A.Compose([...]) snippet matches your exact slider settings, one-click copy
URL sharing	All settings encoded in the URL hash — share your pipeline config with a link
Seed control	Lock a seed for reproducible results or shuffle for exploration
Keyboard shortcuts	Enter to apply, R to shuffle seed

---

Quick start

# Docker — zero setup
git clone https://github.com/YOUR_USERNAME/augmentation-playground
cd augmentation-playground
docker compose up

Open http://localhost:8000

# Or plain Python
pip install -r requirements.txt
python main.py

---

Augmentations

Category	Transforms
Geometric	HorizontalFlip VerticalFlip RandomRotate90 Rotate RandomScale
Color / exposure	RandomBrightnessContrast HueSaturationValue RGBShift CLAHE Equalize
Noise	GaussNoise
Blur	GaussianBlur MotionBlur
Dropout	CoarseDropout (cutout)

---

Label format

Expects YOLO polygon / segmentation format — the same format output by tools like Roboflow, CVAT, and Label Studio:

class_id  x1 y1  x2 y2  x3 y3  ...  xN yN

Coordinates are normalized [0, 1]. Any number of points per grain. Multiple grains per file.

---

REST API

The backend is plain FastAPI — use it directly from your training scripts:

import requests, base64

with open("image.jpg", "rb") as f:
    b64 = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode()

with open("image.txt") as f:
    labels = f.read()

res = requests.post("http://localhost:8000/augment", json={
    "image_b64": b64,
    "label_text": labels,
    "params": {
        "hflip": True,
        "rotate_limit": 30,
        "brightness_limit": 0.4,
        "gauss_noise": 0.05,
        "seed": 42,
    }
})

data = res.json()
# data["aug_clean_b64"]      — augmented image (clean)
# data["aug_annot_b64"]      — augmented image with polygons drawn
# data["aug_polygon_count"]  — how many polygons survived the transform
# data["code"]               — ready-to-paste Albumentations snippet

Endpoint	Description
POST /upload	Upload image file → base64
POST /preview_labels	Draw polygons on original image
POST /augment	Augment single image, returns clean + annotated versions
POST /augment_grid	Augment N variants, returns clean + annotated grids

---

Stack

• Backend: Python · FastAPI · Albumentations · OpenCV — main.py is ~300 lines
• Frontend: single index.html — no React, no build step, no Node
• Deploy: docker compose up

---

Contributing

PRs welcome. Most wanted:

• Live update on slider drag (debounced, no apply button needed)
• Export augmented dataset (apply pipeline to folder of images)
• More transforms (ElasticTransform, GridDistortion, MixUp)
• Drag-and-drop reordering of transforms

---

License

MIT