| |
| |
| """CASID 图块裁切脚本（参考 DOTA 数据集的大图滑窗裁切方式，面向语义分割）。 |
| |
| DOTA（arXiv:1711.10398）将大图按滑动窗口切成 1024×1024 图块（论文默认步距 512）， |
| DOTA_devkit 的 ImgSplit 工具把图块尺寸(subsize)与重叠(gap)做成参数，并支持多尺度 |
| （先把原图按 rate 缩放，再用同样的窗口切）。本脚本沿用这套约定，并针对语义分割 |
| 做了适配：图像与标注逐像素同步裁切，标注按类别 id 处理（缩放一律用最近邻插值）， |
| 窗口越界部分用 0（background）填充或按边缘对齐截取。 |
| |
| 默认参数与 CASID 论文实验完全一致：1024×1024、步距 994（重叠 30）， |
| 每张 5000×5000 原图恰好切出 5×5=25 块，无填充、无浪费。 |
| |
| 依赖：Python >= 3.7，Pillow，numpy |
| |
| 用法示例： |
| # 论文同款（默认，25 块/张） |
| python casid_split.py --root CASID/ --out TILES_1024/ |
| |
| # DOTA 论文设置：重叠 512，边缘补零 |
| python casid_split.py --root CASID/ --out TILES_DOTA/ --gap 512 --pad |
| |
| # mmrotate 社区常用：重叠 200 |
| python casid_split.py --root CASID/ --out TILES_gap200/ --gap 200 --pad |
| |
| # 多尺度（DOTA 多尺度惯例 0.5/1.0/1.5） |
| python casid_split.py --root CASID/ --out TILES_MS/ --rates 0.5 1.0 1.5 |
| |
| # 只切训练集 / 指定区域 |
| python casid_split.py --root CASID/ --out TILES/ --split train |
| python casid_split.py --root CASID/ --out TILES/ --regions 20,21 |
| |
| 输入目录结构（即 Hugging Face 上 CASID 仓库的结构）： |
| CASID/ |
| images/{zone}/{region_dir}/*_img.tif |
| labels/{zone}/{region_dir}/*_gt.png |
| splits/train.txt, test.txt # 每行一个样本 stem（相对路径不含扩展名） |
| regions.csv |
| |
| 输出： |
| out/images/{zone}/{region_dir}/{stem}_H{y}_W{x}.png |
| out/labels/{zone}/{region_dir}/{stem}_H{y}_W{x}_gt.png |
| 多尺度时外层多一级 out/rate{r}/... |
| """ |
| import argparse |
| import csv |
| import os |
| import sys |
| from multiprocessing import Pool |
|
|
| import numpy as np |
| from PIL import Image |
|
|
| |
| CLASS_NAMES = ["background", "building", "forest", "road", "water"] |
| PALETTE_RGB = { |
| 0: (0, 0, 0), |
| 1: (255, 255, 0), |
| 2: (85, 255, 0), |
| 3: (255, 0, 0), |
| 4: (0, 112, 255), |
| } |
| N_CLASSES = 5 |
| PAL_SIZE = 16 |
|
|
|
|
| def build_palette(): |
| pal = [] |
| for i in range(PAL_SIZE): |
| pal.extend(PALETTE_RGB.get(i, (i, i, i))) |
| return pal |
|
|
|
|
| PALETTE = build_palette() |
|
|
|
|
| def tile_starts(size, subsize, stride, pad): |
| """一个轴上的窗口起点序列。 |
| |
| pad=False：固定步距网格（与 CASID 原始图块一致，所有图像共用同一网格； |
| 上一窗未到达边缘才继续开窗，末窗越界时按图像边缘裁齐——如 4656 高的图 |
| 网格仍为 {0,994,1988,2982,3976}，末行高度裁到 680）。 |
| pad=True：DOTA 风格，按 stride 铺满、末窗越界补零。 |
| """ |
| if size <= subsize: |
| return [0] |
| if pad: |
| return list(range(0, size, stride)) |
| starts = [0] |
| while starts[-1] + subsize < size: |
| starts.append(starts[-1] + stride) |
| return starts |
|
|
|
|
| def load_pairs(root, args): |
| """返回 [(img_path, lbl_path, stem, zone, region_dir), ...]""" |
| pairs = [] |
| zones = args.zones.split(",") if args.zones else None |
| regions = set(args.regions.split(",")) if args.regions else None |
| allow = None |
| if args.split and args.split != "all": |
| sf = os.path.join(root, "splits", "%s.txt" % args.split) |
| with open(sf, encoding="utf-8") as f: |
| allow = set(l.strip() for l in f if l.strip()) |
| with open(os.path.join(root, "regions.csv"), encoding="utf-8") as f: |
| for row in csv.DictReader(f): |
| if zones and row["zone_en"] not in zones: |
| continue |
| if regions and row["region_id"] not in regions: |
| continue |
| zone, rdir = row["zone_en"], row["region_dir"] |
| idir = os.path.join(root, "images", zone, rdir) |
| ldir = os.path.join(root, "labels", zone, rdir) |
| if not os.path.isdir(idir): |
| print("WARNING: 找不到 %s，跳过区域 %s（本地未下载该区？）" % (idir, rdir)) |
| continue |
| for fn in sorted(os.listdir(idir)): |
| if not fn.endswith(args.img_ext): |
| continue |
| stem = fn[: -len(args.img_ext)] |
| if stem.endswith("_img"): |
| stem = stem[: -len("_img")] |
| ipath = os.path.join(idir, fn) |
| lpath = os.path.join(ldir, stem + args.lbl_suf + ".png") |
| if not os.path.isfile(lpath): |
| raise FileNotFoundError("label missing for %s" % ipath) |
| if allow is not None: |
| rel = "%s/%s/%s" % (zone, rdir, stem) |
| if rel not in allow: |
| continue |
| pairs.append((ipath, lpath, stem, zone, rdir)) |
| return pairs |
|
|
|
|
| def cut_one(task): |
| ipath, lpath, stem, zone, rdir, out_root, subsize, stride, pad, rate = task |
| img = Image.open(ipath); img.load() |
| lbl = Image.open(lpath); lbl.load() |
| |
| la = np.array(lbl) |
| if la.ndim == 3: |
| la = la[:, :, 0] |
| la = la.astype(np.uint8, copy=False) |
| if rate != 1.0: |
| w, h = img.size |
| nw, nh = round(w * rate), round(h * rate) |
| img = img.resize((nw, nh), Image.BILINEAR) |
| la_img = Image.fromarray(la, "L").resize((nw, nh), Image.NEAREST) |
| la = np.array(la_img) |
| W, H = img.size |
| assert la.shape == (H, W), "label/image size mismatch: %s" % ipath |
| ia = np.asarray(img) |
| xs = tile_starts(W, subsize, stride, pad) |
| ys = tile_starts(H, subsize, stride, pad) |
| od_i = os.path.join(out_root, "images", zone, rdir) |
| od_l = os.path.join(out_root, "labels", zone, rdir) |
| n = 0 |
| for y in ys: |
| for x in xs: |
| if pad: |
| ch = min(subsize, H - y); cw = min(subsize, W - x) |
| iarr = np.zeros((subsize, subsize, 3), dtype=np.uint8) |
| iarr[:ch, :cw] = ia[y:y + ch, x:x + cw] |
| larr = np.zeros((subsize, subsize), dtype=np.uint8) |
| larr[:ch, :cw] = la[y:y + ch, x:x + cw] |
| else: |
| iarr = np.ascontiguousarray(ia[y:y + subsize, x:x + subsize]) |
| larr = np.ascontiguousarray(la[y:y + subsize, x:x + subsize]) |
| itile = Image.fromarray(iarr) |
| ltile = Image.fromarray(larr) |
| ltile.putpalette(PALETTE) |
| os.makedirs(od_i, exist_ok=True) |
| os.makedirs(od_l, exist_ok=True) |
| base = "%s_H%d_W%d" % (stem, y, x) |
| itile.save(os.path.join(od_i, base + ".png")) |
| ltile.save(os.path.join(od_l, base + "_gt.png")) |
| n += 1 |
| return (stem, n) |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) |
| ap.add_argument("--root", required=True, help="CASID 数据集根目录") |
| ap.add_argument("--out", required=True, help="输出根目录") |
| ap.add_argument("--subsize", type=int, default=1024, help="图块边长（默认 1024）") |
| ap.add_argument("--stride", type=int, default=994, help="滑动步距（默认 994，即重叠 30，论文同款）") |
| ap.add_argument("--gap", type=int, default=None, |
| help="DOTA 习惯：用重叠量代替步距，stride = subsize - gap（如 --gap 512 为 DOTA 论文设置）") |
| ap.add_argument("--pad", action="store_true", help="越界补零（DOTA 风格）；默认关闭：末窗锚定图像边缘，不补零") |
| ap.add_argument("--rates", type=float, nargs="+", default=[1.0], help="多尺度缩放系数（DOTA 惯例 0.5 1.0 1.5）") |
| ap.add_argument("--split", default="all", choices=["train", "test", "all"]) |
| ap.add_argument("--zones", default=None, help="逗号分隔的气候带英文名过滤") |
| ap.add_argument("--regions", default=None, help="逗号分隔的区域编号过滤，如 20,21") |
| ap.add_argument("--img-ext", default=".tif") |
| ap.add_argument("--lbl-suf", default="_gt") |
| ap.add_argument("--workers", type=int, default=8) |
| args = ap.parse_args() |
| if args.gap is not None: |
| args.stride = args.subsize - args.gap |
| if args.stride <= 0: |
| sys.exit("--gap 必须小于 --subsize") |
| if args.stride > args.subsize: |
| sys.exit("--stride 不能大于 --subsize（会产生漏切空隙）") |
|
|
| pairs = load_pairs(args.root, args) |
| print("待处理样本：%d" % len(pairs)) |
|
|
| for rate in args.rates: |
| if rate <= 0: |
| sys.exit("--rates 必须为正数") |
| out_root = args.out if len(args.rates) == 1 else os.path.join(args.out, "rate%s" % rate) |
| tasks = [(ip, lp, st, z, r, out_root, args.subsize, args.stride, args.pad, rate) |
| for ip, lp, st, z, r in pairs] |
| total = 0 |
| with Pool(args.workers) as pool: |
| for stem, n in pool.imap_unordered(cut_one, tasks): |
| total += n |
| print("rate=%s：完成 %d 样本，共 %d 图块 → %s" % (rate, len(tasks), total, out_root)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
