target_roi_yolo.py
This commit is contained in:
gcw_4spBpAfv
875
test/synth_compose_yolo.py
Normal file
875
test/synth_compose_yolo.py
Normal file
@@ -0,0 +1,875 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
合成训练数据：把「靶子」贴到随机背景上，并自动生成标注（无需手工标注）。
|
||||
|
||||
前置条件（推荐）：
|
||||
- 靶子用带透明通道的 PNG（抠图后），脚本按非透明像素算紧贴 bbox；
|
||||
- 若只有矩形靶图无 alpha，可用整张图作为矩形框贴入（略松）。
|
||||
|
||||
输出（默认 Pascal VOC，适配 MaixCam 等平台）：
|
||||
- images/xxx.jpg
|
||||
- xml/xxx.xml（与图片同名；单目标或多目标时可扩展）
|
||||
- 生成张数不超过 --max-images（默认 3000）
|
||||
可选 YOLO：
|
||||
- labels/xxx.txt（class cx cy w h，相对 0~1）
|
||||
|
||||
多三角形检测（Pascal VOC 多 <object>，适配 YOLOv5 转 VOC 训练）：
|
||||
- 提供 --triangles-json，顶点在与 --fg 一致的原始靶图像素坐标系下；
|
||||
- 脚本先按 alpha 外接框裁切靶图，顶点会自动减去裁切偏移；
|
||||
- 透视变换时同步变换顶点，每张图输出多个三角形框；
|
||||
- 默认标注为顶点轴对齐最小外接矩形；可选 --triangle-bbox-pad-frac 四周加比例余量（与推理 margin 对齐）。
|
||||
|
||||
Stage2 ROI（对齐「先检整靶再裁小图」的第二步输入）：
|
||||
- --stage2-crop：在合成+增强后，按靶子外接框四周随机 padding 裁剪，标注改到裁剪图坐标系；
|
||||
- 有 --triangles-json 时默认要求裁剪后三角形数与 JSON 一致，否则丢弃重采样（可用 --stage2-allow-partial）。
|
||||
|
||||
运动模糊（模拟手持/快门，默认约一半样本会施加；标注仍为几何真值，与真机域更接近）：
|
||||
- --motion-prob：施加概率；--motion-kernel-min/max：模糊 streak 长度（奇数核，越大越糊）。
|
||||
- 可与 --blur-max 高斯模糊叠加；Stage2 建议：--motion-prob 0.5~0.7 --motion-kernel-max 35 --blur-max 1.2
|
||||
|
||||
依赖：OpenCV + NumPy（PC 上跑即可；Maix 上若内存够也可试）。
|
||||
|
||||
示例：
|
||||
python test/synth_compose_yolo.py --bg-dir ./bg --fg ./target_cutout.png --out ./synth_out --num 3000
|
||||
python test/synth_compose_yolo.py ... --triangles-json test/archery_triangles_default.json --class-name triangle --stage2-crop
|
||||
python test/synth_compose_yolo.py ... --zip ./dataset_voc.zip
|
||||
python test/synth_compose_yolo.py ... --format yolo --out ./synth_yolo
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import random
|
||||
import sys
|
||||
import zipfile
|
||||
import xml.etree.ElementTree as ET
|
||||
|
||||
import numpy as np
|
||||
|
||||
|
||||
def _collect_images(folder: str, exts=(".jpg", ".jpeg", ".png", ".bmp")):
|
||||
out = []
|
||||
for name in sorted(os.listdir(folder)):
|
||||
low = name.lower()
|
||||
if low.endswith(exts):
|
||||
out.append(os.path.join(folder, name))
|
||||
return out
|
||||
|
||||
|
||||
def _load_triangles_json(path: str) -> list[list[tuple[float, float]]]:
|
||||
with open(path, encoding="utf-8") as f:
|
||||
data = json.load(f)
|
||||
tris = data.get("triangles")
|
||||
if not isinstance(tris, list) or not tris:
|
||||
raise ValueError(f'JSON 需包含非空 "triangles" 数组: {path}')
|
||||
out: list[list[tuple[float, float]]] = []
|
||||
for t in tris:
|
||||
if not isinstance(t, list) or len(t) != 3:
|
||||
raise ValueError(f"每个三角形需 3 个顶点: {t!r}")
|
||||
pts = []
|
||||
for p in t:
|
||||
if not isinstance(p, (list, tuple)) or len(p) != 2:
|
||||
raise ValueError(f"顶点需为 [x,y]: {p!r}")
|
||||
pts.append((float(p[0]), float(p[1])))
|
||||
out.append(pts)
|
||||
return out
|
||||
|
||||
|
||||
def _warp_triangle_points(
|
||||
corners_fg_orig: list[tuple[float, float]],
|
||||
fx0: float,
|
||||
fy0: float,
|
||||
fw0: float,
|
||||
fh0: float,
|
||||
new_w: int,
|
||||
new_h: int,
|
||||
persp_M,
|
||||
px: int,
|
||||
py: int,
|
||||
np,
|
||||
cv2,
|
||||
) -> np.ndarray:
|
||||
"""原始靶图像素坐标下的三角形顶点 -> 合成图上的 (3,2) float32。"""
|
||||
pts = np.array(corners_fg_orig, dtype=np.float32)
|
||||
pts[:, 0] -= fx0
|
||||
pts[:, 1] -= fy0
|
||||
pts[:, 0] *= new_w / max(fw0, 1e-6)
|
||||
pts[:, 1] *= new_h / max(fh0, 1e-6)
|
||||
if persp_M is not None:
|
||||
pts = cv2.perspectiveTransform(pts.reshape(1, -1, 2), persp_M).reshape(-1, 2)
|
||||
pts[:, 0] += px
|
||||
pts[:, 1] += py
|
||||
return pts
|
||||
|
||||
|
||||
def _triangle_xyxy_exclusive(
|
||||
pts_xy: np.ndarray, img_w: int, img_h: int
|
||||
) -> tuple[int, int, int, int] | None:
|
||||
xs = pts_xy[:, 0]
|
||||
ys = pts_xy[:, 1]
|
||||
bx0 = max(0, min(img_w - 1, int(np.floor(float(xs.min())))))
|
||||
by0 = max(0, min(img_h - 1, int(np.floor(float(ys.min())))))
|
||||
bx1 = max(bx0 + 1, min(img_w, int(np.ceil(float(xs.max())))))
|
||||
by1 = max(by0 + 1, min(img_h, int(np.ceil(float(ys.max())))))
|
||||
if bx1 <= bx0 or by1 <= by0:
|
||||
return None
|
||||
return bx0, by0, bx1, by1
|
||||
|
||||
|
||||
def _expand_xyxy_half_open(
|
||||
bx0: int,
|
||||
by0: int,
|
||||
bx1: int,
|
||||
by1: int,
|
||||
img_w: int,
|
||||
img_h: int,
|
||||
pad_frac: float,
|
||||
) -> tuple[int, int, int, int] | None:
|
||||
"""在半开框 [bx0,bx1)×[by0,by1) 四周按 max(宽,高)×pad_frac 对称扩展，并裁入图像。"""
|
||||
if pad_frac <= 1e-9:
|
||||
return bx0, by0, bx1, by1
|
||||
bw = max(1, bx1 - bx0)
|
||||
bh = max(1, by1 - by0)
|
||||
base = float(max(bw, bh))
|
||||
p = float(pad_frac) * base
|
||||
x0 = int(np.floor(float(bx0) - p))
|
||||
y0 = int(np.floor(float(by0) - p))
|
||||
x1 = int(np.ceil(float(bx1) + p))
|
||||
y1 = int(np.ceil(float(by1) + p))
|
||||
iw, ih = max(1, img_w), max(1, img_h)
|
||||
x0 = max(0, min(x0, iw - 1))
|
||||
y0 = max(0, min(y0, ih - 1))
|
||||
x1 = max(x0 + 1, min(x1, iw))
|
||||
y1 = max(y0 + 1, min(y1, ih))
|
||||
if x1 <= x0 or y1 <= y0:
|
||||
return None
|
||||
return x0, y0, x1, y1
|
||||
|
||||
|
||||
def _stage2_crop_window(
|
||||
tx0: int,
|
||||
ty0: int,
|
||||
tx1: int,
|
||||
ty1: int,
|
||||
img_w: int,
|
||||
img_h: int,
|
||||
pad_min_frac: float,
|
||||
pad_max_frac: float,
|
||||
rng: random.Random,
|
||||
) -> tuple[int, int, int, int] | None:
|
||||
"""
|
||||
以靶子轴对齐框 [tx0,tx1)×[ty0,ty1)（半开）为中心，四周加随机 padding（相对 max(宽,高) 的比例），
|
||||
再限制在图像内。返回 (cx0, cy0, cw, ch) 用于 comp[cy0:cy0+ch, cx0:cx0+cw]。
|
||||
"""
|
||||
iw, ih = max(1, img_w), max(1, img_h)
|
||||
tw = max(1, tx1 - tx0)
|
||||
th = max(1, ty1 - ty0)
|
||||
base = float(max(tw, th))
|
||||
p0 = max(0.0, float(pad_min_frac))
|
||||
p1 = max(p0, float(pad_max_frac))
|
||||
pad = rng.uniform(p0, p1) * base
|
||||
cx0 = int(np.floor(float(tx0) - pad))
|
||||
cy0 = int(np.floor(float(ty0) - pad))
|
||||
cx1 = int(np.ceil(float(tx1) + pad))
|
||||
cy1 = int(np.ceil(float(ty1) + pad))
|
||||
cx0 = max(0, min(cx0, iw - 1))
|
||||
cy0 = max(0, min(cy0, ih - 1))
|
||||
cx1 = max(cx0 + 1, min(cx1, iw))
|
||||
cy1 = max(cy0 + 1, min(cy1, ih))
|
||||
cw, ch = cx1 - cx0, cy1 - cy0
|
||||
if cw < 4 or ch < 4:
|
||||
return None
|
||||
return cx0, cy0, cw, ch
|
||||
|
||||
|
||||
def _triangle_to_voc_tuple(
|
||||
pts_xy: np.ndarray,
|
||||
img_w: int,
|
||||
img_h: int,
|
||||
class_name: str,
|
||||
bbox_pad_frac: float = 0.0,
|
||||
) -> tuple | None:
|
||||
"""
|
||||
返回 (VOC 元组, 半开 xyxy)；半开框与 VOC 一致地经 pad 扩展，供 YOLO 行写入。
|
||||
bbox_pad_frac>0 时在紧三角形 AABB 四周加 max(宽,高)×frac 余量（truncated 仍按顶点是否贴边）。
|
||||
"""
|
||||
xyxy = _triangle_xyxy_exclusive(pts_xy, img_w, img_h)
|
||||
if xyxy is None:
|
||||
return None
|
||||
bx0, by0, bx1, by1 = xyxy
|
||||
if bbox_pad_frac > 1e-9:
|
||||
exp = _expand_xyxy_half_open(
|
||||
bx0, by0, bx1, by1, img_w, img_h, bbox_pad_frac
|
||||
)
|
||||
if exp is None:
|
||||
return None
|
||||
bx0, by0, bx1, by1 = exp
|
||||
xs = pts_xy[:, 0]
|
||||
ys = pts_xy[:, 1]
|
||||
truncated = (
|
||||
"1"
|
||||
if (
|
||||
xs.min() < -1e-3
|
||||
or xs.max() >= img_w - 1e-3
|
||||
or ys.min() < -1e-3
|
||||
or ys.max() >= img_h - 1e-3
|
||||
)
|
||||
else "0"
|
||||
)
|
||||
vx0, vy0, vx1, vy1 = _xyxy_exclusive_to_voc_inclusive(
|
||||
bx0, by0, bx1, by1, img_w, img_h
|
||||
)
|
||||
if vx1 < vx0 or vy1 < vy0:
|
||||
return None
|
||||
voc = (class_name, vx0, vy0, vx1, vy1, truncated)
|
||||
return voc, (bx0, by0, bx1, by1)
|
||||
|
||||
|
||||
def _fg_bbox_from_alpha(fg_bgra):
|
||||
"""非透明区域的外接矩形 (x,y,w,h)，BGRA。"""
|
||||
import numpy as np
|
||||
|
||||
if fg_bgra.shape[2] < 4:
|
||||
h, w = fg_bgra.shape[:2]
|
||||
return 0, 0, w, h
|
||||
a = fg_bgra[:, :, 3]
|
||||
ys, xs = np.where(a > 10)
|
||||
if len(xs) == 0:
|
||||
h, w = fg_bgra.shape[:2]
|
||||
return 0, 0, w, h
|
||||
x0, x1 = int(xs.min()), int(xs.max())
|
||||
y0, y1 = int(ys.min()), int(ys.max())
|
||||
return x0, y0, x1 - x0 + 1, y1 - y0 + 1
|
||||
|
||||
|
||||
def _paste_fg_on_bg(bg_bgr, x, y, fg_scaled_bgra):
|
||||
"""左上角 (x,y) 将 fg_scaled_bgra（BGRA）贴到 bg_bgr，就地改 bg。"""
|
||||
import numpy as np
|
||||
|
||||
fh, fw = fg_scaled_bgra.shape[:2]
|
||||
bh, bw = bg_bgr.shape[:2]
|
||||
x0, y0 = max(0, x), max(0, y)
|
||||
x1, y1 = min(bw, x + fw), min(bh, y + fh)
|
||||
if x0 >= x1 or y0 >= y1:
|
||||
return
|
||||
fx0, fy0 = x0 - x, y0 - y
|
||||
fx1, fy1 = fx0 + (x1 - x0), fy0 + (y1 - y0)
|
||||
roi_bg = bg_bgr[y0:y1, x0:x1]
|
||||
roi_fg = fg_scaled_bgra[fy0:fy1, fx0:fx1]
|
||||
a = roi_fg[:, :, 3:4].astype(np.float32) / 255.0
|
||||
fg_rgb = roi_fg[:, :, :3].astype(np.float32)
|
||||
bg_rgb = roi_bg.astype(np.float32)
|
||||
blended = fg_rgb * a + bg_rgb * (1.0 - a)
|
||||
roi_bg[:] = blended.astype(np.uint8)
|
||||
|
||||
|
||||
def _perspective_warp_rgba(img_bgra, jitter_frac: float, rng: random.Random, np, cv2):
|
||||
"""
|
||||
对前景做轻微透视（四角微移），返回 (warped BGRA, M)。
|
||||
M 为 3×3，将透视前图像平面上的点映射到 warped 图像像素坐标；未应用透视时返回 (copy, None)。
|
||||
jitter_frac：扰动幅度约为 min(w,h) 的比例。
|
||||
"""
|
||||
h, w = img_bgra.shape[:2]
|
||||
if jitter_frac <= 0 or min(w, h) < 16:
|
||||
return img_bgra.copy(), None
|
||||
|
||||
j = float(max(1.5, min(w, h) * jitter_frac))
|
||||
|
||||
def dj():
|
||||
return rng.uniform(-j, j)
|
||||
|
||||
pts_src = np.float32([[0, 0], [w, 0], [w, h], [0, h]])
|
||||
pts_dst = np.float32(
|
||||
[
|
||||
[dj(), dj()],
|
||||
[w + dj(), dj()],
|
||||
[w + dj(), h + dj()],
|
||||
[dj(), h + dj()],
|
||||
]
|
||||
)
|
||||
|
||||
xmin = float(pts_dst[:, 0].min())
|
||||
ymin = float(pts_dst[:, 1].min())
|
||||
pts_shift = pts_dst.copy()
|
||||
pts_shift[:, 0] -= xmin
|
||||
pts_shift[:, 1] -= ymin
|
||||
out_w = max(4, int(np.ceil(float(pts_shift[:, 0].max()))) + 2)
|
||||
out_h = max(4, int(np.ceil(float(pts_shift[:, 1].max()))) + 2)
|
||||
|
||||
M = cv2.getPerspectiveTransform(pts_src, pts_shift)
|
||||
warped = cv2.warpPerspective(
|
||||
img_bgra,
|
||||
M,
|
||||
(out_w, out_h),
|
||||
flags=cv2.INTER_LINEAR,
|
||||
borderMode=cv2.BORDER_CONSTANT,
|
||||
borderValue=(0, 0, 0, 0),
|
||||
)
|
||||
return warped, M
|
||||
|
||||
|
||||
def _color_jitter_bgr(comp_bgr, strength: float, rng: random.Random, np, cv2):
|
||||
"""整图 HSV 抖动：strength∈[0,1] 越大越强。"""
|
||||
if strength <= 1e-6:
|
||||
return comp_bgr
|
||||
strength = min(1.0, max(0.0, strength))
|
||||
hsv = cv2.cvtColor(comp_bgr, cv2.COLOR_BGR2HSV).astype(np.float32)
|
||||
dh = rng.uniform(-18.0 * strength, 18.0 * strength)
|
||||
hsv[:, :, 0] = (hsv[:, :, 0] + dh) % 180.0
|
||||
sf = rng.uniform(1.0 - 0.22 * strength, 1.0 + 0.22 * strength)
|
||||
vf = rng.uniform(1.0 - 0.22 * strength, 1.0 + 0.22 * strength)
|
||||
hsv[:, :, 1] = np.clip(hsv[:, :, 1] * sf, 0, 255)
|
||||
hsv[:, :, 2] = np.clip(hsv[:, :, 2] * vf, 0, 255)
|
||||
# 轻微 BGR 通道偏置（模拟白平衡）
|
||||
out = cv2.cvtColor(hsv.astype(np.uint8), cv2.COLOR_HSV2BGR).astype(np.float32)
|
||||
bias = np.array(
|
||||
[
|
||||
rng.uniform(-12 * strength, 12 * strength),
|
||||
rng.uniform(-12 * strength, 12 * strength),
|
||||
rng.uniform(-12 * strength, 12 * strength),
|
||||
],
|
||||
dtype=np.float32,
|
||||
)
|
||||
out = np.clip(out + bias, 0, 255).astype(np.uint8)
|
||||
return out
|
||||
|
||||
|
||||
def _motion_blur_bgr(
|
||||
comp_bgr,
|
||||
rng: random.Random,
|
||||
k_min: int,
|
||||
k_max: int,
|
||||
np,
|
||||
cv2,
|
||||
):
|
||||
"""
|
||||
方向随机的线性运动模糊（filter2D）。核为奇数 k×k，沿穿过中心、角度 uniform[0,180°) 的线段归一化求和。
|
||||
标注无需改：bbox 仍为物体真实位置，与真实相机「糊图+真框」的训练惯例一致。
|
||||
"""
|
||||
lo = int(max(3, k_min | 1))
|
||||
hi = int(max(lo, k_max | 1))
|
||||
k = rng.randint(lo, hi)
|
||||
if k % 2 == 0:
|
||||
k = min(hi, k + 1)
|
||||
k = max(3, k)
|
||||
ker_u = np.zeros((k, k), dtype=np.uint8)
|
||||
ang = rng.uniform(0.0, 180.0)
|
||||
rad = float(np.deg2rad(ang))
|
||||
c = k // 2
|
||||
dx = float(np.cos(rad) * (k // 2))
|
||||
dy = float(np.sin(rad) * (k // 2))
|
||||
x0 = int(round(c - dx))
|
||||
y0 = int(round(c - dy))
|
||||
x1 = int(round(c + dx))
|
||||
y1 = int(round(c + dy))
|
||||
cv2.line(ker_u, (x0, y0), (x1, y1), 255, 1)
|
||||
s = float(ker_u.sum())
|
||||
if s < 1e-3:
|
||||
ker_u[c, c] = 255
|
||||
s = 255.0
|
||||
ker = ker_u.astype(np.float32) / s
|
||||
return cv2.filter2D(comp_bgr, -1, ker)
|
||||
|
||||
|
||||
def _yolo_line(cls: int, xyxy_on_bg, img_w: int, img_h: int) -> str:
|
||||
x0, y0, x1, y1 = xyxy_on_bg
|
||||
bw, bh = x1 - x0, y1 - y0
|
||||
cx = (x0 + x1) / 2.0 / img_w
|
||||
cy = (y0 + y1) / 2.0 / img_h
|
||||
nw = bw / img_w
|
||||
nh = bh / img_h
|
||||
cx = max(0.0, min(1.0, cx))
|
||||
cy = max(0.0, min(1.0, cy))
|
||||
nw = max(1e-6, min(1.0, nw))
|
||||
nh = max(1e-6, min(1.0, nh))
|
||||
return f"{cls} {cx:.6f} {cy:.6f} {nw:.6f} {nh:.6f}\n"
|
||||
|
||||
|
||||
def _xyxy_exclusive_to_voc_inclusive(
|
||||
x0: float, y0: float, x1: float, y1: float, img_w: int, img_h: int
|
||||
) -> tuple[int, int, int, int]:
|
||||
"""内部 xyxy 为半开区间 [x0,x1)×[y0,y1)，转为 VOC inclusive 整数像素框。"""
|
||||
iw, ih = max(1, img_w), max(1, img_h)
|
||||
xi0 = max(0, min(iw - 1, int(x0)))
|
||||
yi0 = max(0, min(ih - 1, int(y0)))
|
||||
xi1 = max(xi0, min(iw - 1, int(x1) - 1))
|
||||
yi1 = max(yi0, min(ih - 1, int(y1) - 1))
|
||||
return xi0, yi0, xi1, yi1
|
||||
|
||||
|
||||
def _write_pascal_voc_xml(
|
||||
xml_path: str,
|
||||
img_filename: str,
|
||||
img_folder: str,
|
||||
img_w: int,
|
||||
img_h: int,
|
||||
depth: int,
|
||||
objects: list[tuple],
|
||||
) -> None:
|
||||
"""
|
||||
objects 每项为 (class_name, xmin, ymin, xmax, ymax) 或
|
||||
(class_name, xmin, ymin, xmax, ymax, truncated)，坐标均为 inclusive 整数像素；
|
||||
truncated 为 \"0\" 或 \"1\"（省略时默认为 \"0\"）。
|
||||
"""
|
||||
root = ET.Element("annotation")
|
||||
ET.SubElement(root, "folder").text = img_folder
|
||||
ET.SubElement(root, "filename").text = img_filename
|
||||
src = ET.SubElement(root, "source")
|
||||
ET.SubElement(src, "database").text = "synthetic_archery"
|
||||
ET.SubElement(src, "annotation").text = "Pascal VOC compatible"
|
||||
sz = ET.SubElement(root, "size")
|
||||
ET.SubElement(sz, "width").text = str(img_w)
|
||||
ET.SubElement(sz, "height").text = str(img_h)
|
||||
ET.SubElement(sz, "depth").text = str(depth)
|
||||
ET.SubElement(root, "segmented").text = "0"
|
||||
for item in objects:
|
||||
if len(item) == 6:
|
||||
name, xmin, ymin, xmax, ymax, truncated = item
|
||||
else:
|
||||
name, xmin, ymin, xmax, ymax = item
|
||||
truncated = "0"
|
||||
obj = ET.SubElement(root, "object")
|
||||
ET.SubElement(obj, "name").text = name
|
||||
ET.SubElement(obj, "pose").text = "Unspecified"
|
||||
ET.SubElement(obj, "truncated").text = str(truncated)
|
||||
ET.SubElement(obj, "difficult").text = "0"
|
||||
bb = ET.SubElement(obj, "bndbox")
|
||||
ET.SubElement(bb, "xmin").text = str(xmin)
|
||||
ET.SubElement(bb, "ymin").text = str(ymin)
|
||||
ET.SubElement(bb, "xmax").text = str(xmax)
|
||||
ET.SubElement(bb, "ymax").text = str(ymax)
|
||||
|
||||
tree = ET.ElementTree(root)
|
||||
try:
|
||||
ET.indent(tree, space=" ")
|
||||
except AttributeError:
|
||||
pass
|
||||
tree.write(xml_path, encoding="utf-8", xml_declaration=True)
|
||||
|
||||
|
||||
def _zip_images_xml(dataset_root: str, zip_path: str) -> None:
|
||||
"""打包 dataset_root 下的 images/ 与 xml/ 到 zip（根目录含这两个文件夹）。"""
|
||||
img_dir = os.path.join(dataset_root, "images")
|
||||
xml_dir = os.path.join(dataset_root, "xml")
|
||||
if not os.path.isdir(img_dir) or not os.path.isdir(xml_dir):
|
||||
raise FileNotFoundError(f"需要存在目录: {img_dir} 与 {xml_dir}")
|
||||
zip_path = os.path.abspath(zip_path)
|
||||
os.makedirs(os.path.dirname(zip_path) or ".", exist_ok=True)
|
||||
with zipfile.ZipFile(zip_path, "w", compression=zipfile.ZIP_DEFLATED) as zf:
|
||||
for folder, arc_prefix in ((img_dir, "images"), (xml_dir, "xml")):
|
||||
for name in sorted(os.listdir(folder)):
|
||||
fp = os.path.join(folder, name)
|
||||
if os.path.isfile(fp):
|
||||
zf.write(fp, arcname=os.path.join(arc_prefix, name).replace("\\", "/"))
|
||||
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--bg-dir", required=True, help="背景图目录")
|
||||
ap.add_argument("--fg", required=True, help="靶子 PNG（推荐 RGBA 抠图）或任意图")
|
||||
ap.add_argument("--out", default="./synth_dataset", help="输出根目录")
|
||||
ap.add_argument("--num", type=int, default=200, help="请求生成张数（实际不超过 --max-images）")
|
||||
ap.add_argument(
|
||||
"--max-images",
|
||||
type=int,
|
||||
default=3000,
|
||||
help="最多生成图片张数，超出部分忽略（MaixCam 等平台常见上限 3000）",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--format",
|
||||
choices=("voc", "yolo", "both"),
|
||||
default="voc",
|
||||
help="voc=Pascal VOC（images+xml）；yolo=labels txt；both=两者都写",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--class-name",
|
||||
default="黑三角和圆环",
|
||||
help="VOC <object><name> 类别名（单类检测默认 target）",
|
||||
)
|
||||
ap.add_argument("--class-id", type=int, default=0, help="YOLO 类别 id（仅 --format yolo/both）")
|
||||
ap.add_argument(
|
||||
"--zip",
|
||||
default=None,
|
||||
metavar="PATH",
|
||||
help="完成后将 images/ 与 xml/ 打成 zip（仅 VOC/both 时有 xml；路径如 ./dataset.zip）",
|
||||
)
|
||||
ap.add_argument("--seed", type=int, default=None)
|
||||
ap.add_argument("--scale-min", type=float, default=0.15, help="靶子最短边占背景最短边比例下限")
|
||||
ap.add_argument("--scale-max", type=float, default=0.55, help="比例上限")
|
||||
ap.add_argument("--blur-max", type=float, default=0.0, help="高斯模糊 sigma 上限，0 关闭")
|
||||
ap.add_argument(
|
||||
"--motion-prob",
|
||||
type=float,
|
||||
default=0.45,
|
||||
help="运动模糊概率 0~1（默认约一半样本；关模糊用 0）",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--motion-kernel-min",
|
||||
type=int,
|
||||
default=7,
|
||||
help="运动模糊 streak 长度下限（奇数，实际会纠到奇数）",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--motion-kernel-max",
|
||||
type=int,
|
||||
default=35,
|
||||
help="运动模糊 streak 长度上限，越大越像长曝光/手抖",
|
||||
)
|
||||
ap.add_argument("--jpeg-quality", type=int, default=92)
|
||||
ap.add_argument(
|
||||
"--perspective",
|
||||
type=float,
|
||||
default=0.0,
|
||||
help="轻微透视：四角扰动约为 min(靶宽,靶高)×该系数，0 关闭（建议 0.02~0.06）",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--perspective-prob",
|
||||
type=float,
|
||||
default=0.75,
|
||||
help="每张图应用透视的概率 0~1",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--color-jitter",
|
||||
type=float,
|
||||
default=0.0,
|
||||
help="合成后整图颜色抖动强度 0~1，0 关闭（建议 0.4~0.8）",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--triangles-json",
|
||||
default=None,
|
||||
metavar="PATH",
|
||||
help="三角形顶点 JSON（test/archery_triangles_default.json）；坐标与 --fg 原图一致，"
|
||||
"多三角形时每张图写多个 VOC <object>（透视时顶点同步变换）",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--triangle-bbox-pad-frac",
|
||||
type=float,
|
||||
default=0.0,
|
||||
help="三角形检测框在紧 AABB 四周再加 max(宽,高)×该比例（VOC/YOLO 同步）；"
|
||||
"0=贴顶点外接框；Stage2 建议 0.08~0.18，与推理端 margin 接近更易对齐",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--stage2-crop",
|
||||
action="store_true",
|
||||
help="合成与增强后按靶子外接框+随机边距裁剪，输出与 Stage2（整靶 ROI）构图一致；标注为裁剪后坐标",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--stage2-pad-min",
|
||||
type=float,
|
||||
default=0.02,
|
||||
help="Stage2 裁剪：四边 padding 相对靶 max(宽,高) 的比例下限",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--stage2-pad-max",
|
||||
type=float,
|
||||
default=0.14,
|
||||
help="Stage2 裁剪：padding 比例上限",
|
||||
)
|
||||
ap.add_argument(
|
||||
"--stage2-allow-partial",
|
||||
action="store_true",
|
||||
help="有 --triangles-json 时允许裁剪后有效三角形数少于 JSON（默认要求数量一致）",
|
||||
)
|
||||
args = ap.parse_args()
|
||||
|
||||
try:
|
||||
import cv2
|
||||
import numpy as np
|
||||
except ImportError:
|
||||
print("[ERR] 需要 opencv-python、numpy")
|
||||
sys.exit(1)
|
||||
|
||||
rng = random.Random(args.seed)
|
||||
|
||||
bgs = _collect_images(args.bg_dir)
|
||||
if not bgs:
|
||||
print(f"[ERR] 背景目录无图片: {args.bg_dir}")
|
||||
sys.exit(1)
|
||||
|
||||
fg_path = args.fg
|
||||
if not os.path.isfile(fg_path):
|
||||
print(f"[ERR] 找不到靶图: {fg_path}")
|
||||
sys.exit(1)
|
||||
|
||||
fg = cv2.imread(fg_path, cv2.IMREAD_UNCHANGED)
|
||||
if fg is None:
|
||||
print(f"[ERR] 无法读取靶图: {fg_path}")
|
||||
sys.exit(1)
|
||||
if fg.ndim == 2:
|
||||
fg = cv2.cvtColor(fg, cv2.COLOR_GRAY2BGRA)
|
||||
elif fg.shape[2] == 3:
|
||||
b, g, r = cv2.split(fg)
|
||||
a = np.full_like(b, 255)
|
||||
fg = cv2.merge([b, g, r, a])
|
||||
|
||||
fx0, fy0, fw0, fh0 = _fg_bbox_from_alpha(fg)
|
||||
fg_crop = fg[fy0 : fy0 + fh0, fx0 : fx0 + fw0].copy()
|
||||
|
||||
triangles_full = None
|
||||
if args.triangles_json:
|
||||
tpath = args.triangles_json
|
||||
if not os.path.isfile(tpath):
|
||||
print(f"[ERR] 找不到 --triangles-json: {tpath}")
|
||||
sys.exit(1)
|
||||
try:
|
||||
triangles_full = _load_triangles_json(tpath)
|
||||
except (json.JSONDecodeError, ValueError, OSError) as e:
|
||||
print(f"[ERR] 解析三角形 JSON 失败: {e}")
|
||||
sys.exit(1)
|
||||
print(f"[INFO] 已加载 {len(triangles_full)} 个三角形（每张图多个 VOC 检测框）")
|
||||
|
||||
want_voc = args.format in ("voc", "both")
|
||||
want_yolo = args.format in ("yolo", "both")
|
||||
n_gen = min(max(0, args.num), max(0, args.max_images))
|
||||
if args.num > args.max_images:
|
||||
print(f"[INFO] --num={args.num} 大于 --max-images={args.max_images}，仅生成 {n_gen} 张")
|
||||
|
||||
if args.stage2_crop:
|
||||
print(
|
||||
f"[INFO] Stage2 裁剪: pad∈[{args.stage2_pad_min},{args.stage2_pad_max}]×max(靶宽,靶高)，"
|
||||
f"partial={'允许' if args.stage2_allow_partial else '不允许'}"
|
||||
)
|
||||
|
||||
out_img = os.path.join(args.out, "images")
|
||||
out_xml = os.path.join(args.out, "xml")
|
||||
out_lbl = os.path.join(args.out, "labels")
|
||||
os.makedirs(out_img, exist_ok=True)
|
||||
if want_voc:
|
||||
os.makedirs(out_xml, exist_ok=True)
|
||||
if want_yolo:
|
||||
os.makedirs(out_lbl, exist_ok=True)
|
||||
|
||||
print(f"[INFO] 背景 {len(bgs)} 张，格式={args.format}，生成 {n_gen} 张 → {args.out}")
|
||||
|
||||
i_done = 0
|
||||
while i_done < n_gen:
|
||||
bg_path = rng.choice(bgs)
|
||||
bg = cv2.imread(bg_path, cv2.IMREAD_COLOR)
|
||||
if bg is None:
|
||||
continue
|
||||
bh, bw = bg.shape[:2]
|
||||
short_bg = min(bh, bw)
|
||||
short_fg = min(fh0, fw0)
|
||||
smin = args.scale_min * short_bg / max(short_fg, 1)
|
||||
smax = args.scale_max * short_bg / max(short_fg, 1)
|
||||
scale = rng.uniform(max(smin, 0.05), max(smax, smin + 0.01))
|
||||
|
||||
new_w = max(4, int(fw0 * scale))
|
||||
new_h = max(4, int(fh0 * scale))
|
||||
fg_s = cv2.resize(fg_crop, (new_w, new_h), interpolation=cv2.INTER_AREA)
|
||||
|
||||
persp_M = None
|
||||
if args.perspective > 0 and rng.random() < args.perspective_prob:
|
||||
fg_s, persp_M = _perspective_warp_rgba(fg_s, args.perspective, rng, np, cv2)
|
||||
|
||||
fw2, fh2 = fg_s.shape[1], fg_s.shape[0]
|
||||
tx0, ty0, tw, th = _fg_bbox_from_alpha(fg_s)
|
||||
|
||||
max_x = max(0, bw - fw2)
|
||||
max_y = max(0, bh - fh2)
|
||||
px = rng.randint(0, max_x) if max_x > 0 else 0
|
||||
py = rng.randint(0, max_y) if max_y > 0 else 0
|
||||
|
||||
comp = bg.copy()
|
||||
_paste_fg_on_bg(comp, px, py, fg_s)
|
||||
|
||||
# 标注：整靶 alpha 框（无 triangles-json 时使用）或多三角形框
|
||||
bx0 = px + tx0
|
||||
by0 = py + ty0
|
||||
bx1 = px + tx0 + tw
|
||||
by1 = py + ty0 + th
|
||||
bx0 = max(0, min(bx0, bw - 1))
|
||||
by0 = max(0, min(by0, bh - 1))
|
||||
bx1 = max(bx0 + 1, min(bx1, bw))
|
||||
by1 = max(by0 + 1, min(by1, bh))
|
||||
|
||||
tri_pts_full: list[np.ndarray] = []
|
||||
if triangles_full is not None:
|
||||
for tri in triangles_full:
|
||||
pts_c = _warp_triangle_points(
|
||||
tri,
|
||||
float(fx0),
|
||||
float(fy0),
|
||||
float(fw0),
|
||||
float(fh0),
|
||||
new_w,
|
||||
new_h,
|
||||
persp_M,
|
||||
px,
|
||||
py,
|
||||
np,
|
||||
cv2,
|
||||
)
|
||||
tri_pts_full.append(pts_c)
|
||||
|
||||
if args.color_jitter > 1e-6:
|
||||
comp = _color_jitter_bgr(comp, args.color_jitter, rng, np, cv2)
|
||||
|
||||
if args.blur_max > 1e-6:
|
||||
sig = rng.uniform(0.3, args.blur_max)
|
||||
k = int(sig * 4) | 1
|
||||
comp = cv2.GaussianBlur(comp, (k, k), sig)
|
||||
|
||||
if rng.random() < max(0.0, min(1.0, float(args.motion_prob))):
|
||||
comp = _motion_blur_bgr(
|
||||
comp,
|
||||
rng,
|
||||
args.motion_kernel_min,
|
||||
args.motion_kernel_max,
|
||||
np,
|
||||
cv2,
|
||||
)
|
||||
|
||||
bh, bw = comp.shape[:2]
|
||||
|
||||
if args.stage2_crop:
|
||||
win = _stage2_crop_window(
|
||||
bx0,
|
||||
by0,
|
||||
bx1,
|
||||
by1,
|
||||
bw,
|
||||
bh,
|
||||
args.stage2_pad_min,
|
||||
args.stage2_pad_max,
|
||||
rng,
|
||||
)
|
||||
if win is None:
|
||||
continue
|
||||
cx0, cy0, cw, ch = win
|
||||
comp = comp[cy0 : cy0 + ch, cx0 : cx0 + cw].copy()
|
||||
out_w, out_h = cw, ch
|
||||
|
||||
if triangles_full is not None:
|
||||
voc_objects = []
|
||||
yolo_lines_list = []
|
||||
for pts_c in tri_pts_full:
|
||||
p2 = pts_c.copy()
|
||||
p2[:, 0] -= cx0
|
||||
p2[:, 1] -= cy0
|
||||
pair = _triangle_to_voc_tuple(
|
||||
p2,
|
||||
out_w,
|
||||
out_h,
|
||||
args.class_name,
|
||||
args.triangle_bbox_pad_frac,
|
||||
)
|
||||
if pair is None:
|
||||
continue
|
||||
vo, xyxy = pair
|
||||
voc_objects.append(vo)
|
||||
if want_yolo:
|
||||
yolo_lines_list.append(
|
||||
_yolo_line(args.class_id, xyxy, out_w, out_h)
|
||||
)
|
||||
if not args.stage2_allow_partial and len(voc_objects) != len(
|
||||
triangles_full
|
||||
):
|
||||
continue
|
||||
if want_voc and not voc_objects:
|
||||
continue
|
||||
if want_yolo and not yolo_lines_list:
|
||||
continue
|
||||
else:
|
||||
nbx0, nby0 = bx0 - cx0, by0 - cy0
|
||||
nbx1, nby1 = bx1 - cx0, by1 - cy0
|
||||
nbx0 = max(0, min(nbx0, out_w - 1))
|
||||
nby0 = max(0, min(nby0, out_h - 1))
|
||||
nbx1 = max(nbx0 + 1, min(nbx1, out_w))
|
||||
nby1 = max(nby0 + 1, min(nby1, out_h))
|
||||
if nbx1 <= nbx0 or nby1 <= nby0:
|
||||
continue
|
||||
vx0, vy0, vx1, vy1 = _xyxy_exclusive_to_voc_inclusive(
|
||||
nbx0, nby0, nbx1, nby1, out_w, out_h
|
||||
)
|
||||
voc_objects = [(args.class_name, vx0, vy0, vx1, vy1)]
|
||||
yolo_lines_list = (
|
||||
[_yolo_line(args.class_id, (nbx0, nby0, nbx1, nby1), out_w, out_h)]
|
||||
if want_yolo
|
||||
else []
|
||||
)
|
||||
else:
|
||||
out_w, out_h = bw, bh
|
||||
if triangles_full is not None:
|
||||
voc_objects = []
|
||||
yolo_lines_list = []
|
||||
for pts_c in tri_pts_full:
|
||||
pair = _triangle_to_voc_tuple(
|
||||
pts_c,
|
||||
out_w,
|
||||
out_h,
|
||||
args.class_name,
|
||||
args.triangle_bbox_pad_frac,
|
||||
)
|
||||
if pair is None:
|
||||
continue
|
||||
vo, xyxy = pair
|
||||
voc_objects.append(vo)
|
||||
if want_yolo:
|
||||
yolo_lines_list.append(
|
||||
_yolo_line(args.class_id, xyxy, out_w, out_h)
|
||||
)
|
||||
if want_voc and not voc_objects:
|
||||
continue
|
||||
if want_yolo and not yolo_lines_list:
|
||||
continue
|
||||
else:
|
||||
vx0, vy0, vx1, vy1 = _xyxy_exclusive_to_voc_inclusive(
|
||||
bx0, by0, bx1, by1, out_w, out_h
|
||||
)
|
||||
voc_objects = [(args.class_name, vx0, vy0, vx1, vy1)]
|
||||
yolo_lines_list = (
|
||||
[_yolo_line(args.class_id, (bx0, by0, bx1, by1), out_w, out_h)]
|
||||
if want_yolo
|
||||
else []
|
||||
)
|
||||
|
||||
stem = f"synth_{i_done:06d}"
|
||||
img_name = stem + ".jpg"
|
||||
img_path = os.path.join(out_img, img_name)
|
||||
cv2.imwrite(img_path, comp, [int(cv2.IMWRITE_JPEG_QUALITY), args.jpeg_quality])
|
||||
|
||||
if want_voc:
|
||||
xml_path = os.path.join(out_xml, stem + ".xml")
|
||||
_write_pascal_voc_xml(
|
||||
xml_path,
|
||||
img_filename=img_name,
|
||||
img_folder="images",
|
||||
img_w=out_w,
|
||||
img_h=out_h,
|
||||
depth=3,
|
||||
objects=voc_objects,
|
||||
)
|
||||
if want_yolo:
|
||||
lbl_path = os.path.join(out_lbl, stem + ".txt")
|
||||
with open(lbl_path, "w", encoding="utf-8") as f:
|
||||
f.writelines(yolo_lines_list)
|
||||
|
||||
i_done += 1
|
||||
if i_done % 50 == 0:
|
||||
print(f" ... {i_done}/{n_gen}")
|
||||
|
||||
parts = [out_img]
|
||||
if want_voc:
|
||||
parts.append(out_xml)
|
||||
if want_yolo:
|
||||
parts.append(out_lbl)
|
||||
print(f"[OK] 完成: " + " , ".join(parts))
|
||||
|
||||
if args.zip:
|
||||
if not want_voc:
|
||||
print("[WARN] --zip 需要 VOC 标注目录 xml/，当前格式未生成 xml，跳过打包")
|
||||
else:
|
||||
try:
|
||||
_zip_images_xml(args.out, args.zip)
|
||||
print(f"[OK] 已打包: {os.path.abspath(args.zip)}")
|
||||
except OSError as e:
|
||||
print(f"[ERR] 打包失败: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user