"""Photo-condition augmentations for the retrieval benchmark. A query is a catalog product shot degraded the way a garment looks when cropped out of a real photo: off-centre framing, tilt, lighting and colour cast, blur, compression and, in the "wild" setting, a non-white background and a partial occlusion (an arm, hair, a bag strap). """ from __future__ import annotations import io import numpy as np from PIL import Image, ImageEnhance, ImageFilter def _random_resized_crop(im: Image.Image, rng: np.random.Generator) -> Image.Image: w, h = im.size area = rng.uniform(0.72, 0.95) * w * h ratio = np.exp(rng.uniform(np.log(0.8), np.log(1.25))) cw = int(min(w, round(np.sqrt(area * ratio)))) ch = int(min(h, round(np.sqrt(area / ratio)))) x = int(rng.integers(0, w - cw + 1)) y = int(rng.integers(0, h - ch + 1)) return im.crop((x, y, x + cw, y + ch)) def _background(size: tuple[int, int], rng: np.random.Generator) -> np.ndarray: w, h = size c1 = rng.uniform(30, 225, size=3) c2 = rng.uniform(30, 225, size=3) t = np.linspace(0, 1, h)[:, None, None] if rng.random() < 0.5 else np.linspace(0, 1, w)[None, :, None] grad = c1 * (1 - t) + c2 * t grad = np.broadcast_to(grad, (h, w, 3)).copy() grad += rng.normal(0, rng.uniform(2, 14), size=(h, w, 3)) return grad.clip(0, 255) def augment(im: Image.Image, rng: np.random.Generator, wild: bool = False) -> Image.Image: im = im.convert("RGB") if wild: arr = np.asarray(im, dtype=np.float32) white = (arr > 236).all(axis=2) bg = _background(im.size, rng) arr[white] = bg[white] im = Image.fromarray(arr.astype(np.uint8)) im = _random_resized_crop(im, rng) angle = float(rng.uniform(-12, 12)) fill = tuple(int(v) for v in np.asarray(im).reshape(-1, 3)[0]) if wild else (255, 255, 255) im = im.rotate(angle, resample=Image.BICUBIC, expand=False, fillcolor=fill) im = ImageEnhance.Brightness(im).enhance(float(rng.uniform(0.75, 1.25))) im = ImageEnhance.Contrast(im).enhance(float(rng.uniform(0.75, 1.25))) im = ImageEnhance.Color(im).enhance(float(rng.uniform(0.7, 1.3))) # Warm/cool colour cast (indoor vs daylight white balance). cast = np.array([1 + rng.uniform(-0.08, 0.08), 1.0, 1 + rng.uniform(-0.08, 0.08)], dtype=np.float32) im = Image.fromarray((np.asarray(im, dtype=np.float32) * cast).clip(0, 255).astype(np.uint8)) if wild: w, h = im.size ow = int(w * rng.uniform(0.25, 0.45)) oh = int(h * rng.uniform(0.25, 0.45)) x, y = int(rng.integers(0, w - ow + 1)), int(rng.integers(0, h - oh + 1)) colour = tuple(int(c) for c in rng.uniform(20, 235, size=3)) occ = Image.new("RGB", (ow, oh), colour) im.paste(occ, (x, y)) blur = float(rng.uniform(0, 1.2)) if blur > 0.2: im = im.filter(ImageFilter.GaussianBlur(blur)) side = int(rng.integers(160, 257)) im.thumbnail((side, side), Image.BILINEAR) buf = io.BytesIO() im.save(buf, format="JPEG", quality=int(rng.integers(35, 76))) return Image.open(io.BytesIO(buf.getvalue())).convert("RGB")