deepellipse-circleid / code /resnet_point_feature.py
qinxunhui's picture
Add small artifacts (code, outputs, configs, LICENSE)
519ab98 verified
Raw History Blame Contribute Delete
18.2 kB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
@Time : 2026/4/3 11:44
@File : resnet_point_feature.py
@Description : 基于 ResNet 特征图位置的深度特征提取;支持单张/批量多图多点采样。
"""
from __future__ import annotations
import argparse
import os
import random
from pathlib import Path
from typing import List, Optional, Sequence, Tuple, Union
import cv2
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import models, transforms
from elipse_feature import (
DEFAULT_IMG_SIZE,
DEFAULT_N_THETA,
EllipseFeatureBundle,
extract_ellipse_features,
preprocess_rgb_to_skeleton,
)
PathLike = Union[str, Path, np.ndarray]
RESNET_DEEP_RAW_DIM = 448 # layer1(64) + layer2(128) + layer3(256)
RESNET_CHECKPOINT_WINDOWS = "model/model_epoch_60_image.pth"
RESNET_CHECKPOINT_LINUX = "checkpoints/model_resnet/resnet_rgb_0.49_0.48.pth"
def default_resnet_checkpoint() -> str:
"""§4.1.3 ResNet 权重:Windows 用 model/,Linux 等用 checkpoints/model_resnet/。"""
if os.name == "nt":
return RESNET_CHECKPOINT_WINDOWS
return RESNET_CHECKPOINT_LINUX
DEFAULT_CHECKPOINT = Path(default_resnet_checkpoint())
IMAGENET_MEAN = [0.485, 0.456, 0.406]
IMAGENET_STD = [0.229, 0.224, 0.225]
def _default_device() -> torch.device:
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
def _rgb_bgr_to_tensor(
rgb_bgr: np.ndarray,
device: Optional[torch.device] = None,
) -> Tuple[torch.Tensor, int, int]:
"""BGR uint8 图像 → 归一化 Tensor [1, 3, H, W]。"""
device = device or _default_device()
rgb = cv2.cvtColor(rgb_bgr, cv2.COLOR_BGR2RGB)
orig_h, orig_w = rgb.shape[:2]
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.ToTensor(),
transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
])
tensor = transform(rgb).unsqueeze(0).to(device)
return tensor, orig_h, orig_w
def _batch_rgb_bgr_to_tensor(
rgb_bgr_list: Sequence[np.ndarray],
device: Optional[torch.device] = None,
) -> torch.Tensor:
"""多张 BGR 图像(需同尺寸)→ [B, 3, H, W]。"""
device = device or _default_device()
tensors = [_rgb_bgr_to_tensor(img, device)[0] for img in rgb_bgr_list]
return torch.cat(tensors, dim=0)
def keypoints_from_ellipse_bundle(bundle: EllipseFeatureBundle) -> np.ndarray:
"""从椭圆特征 bundle 取 ResNet 采样坐标 s_base_px,形状 (N_theta, 2),顺序 (x, y)。"""
return bundle.sequence.s_base_px.astype(np.float32)
def extract_ellipse_keypoints_from_bgr(
rgb_bgr: np.ndarray,
img_size: int = DEFAULT_IMG_SIZE,
n_theta: int = DEFAULT_N_THETA,
min_valid: int = 180,
) -> Optional[Tuple[np.ndarray, np.ndarray, EllipseFeatureBundle]]:
"""
BGR 原图 → 骨架化 → 椭圆序列 → 关键点坐标。
Returns
-------
rgb_resized, keypoints_xy, bundle
rgb_resized: resize 后的 BGR,与 keypoints 坐标系一致
keypoints_xy: (N_theta, 2),像素坐标 (x, y)
"""
prep = preprocess_rgb_to_skeleton(rgb_bgr, size=img_size)
bundle = extract_ellipse_features(
prep["skeleton"],
n_theta=n_theta,
half_norm=img_size / 2.0,
compute_fft=False,
)
valid_count = int(np.sum(bundle.sequence.interp_flag == 0))
if valid_count < min_valid:
return None
return prep["rgb"], keypoints_from_ellipse_bundle(bundle), bundle
def extract_ellipse_keypoints_from_path(
image_path: PathLike,
**kwargs,
) -> Optional[Tuple[np.ndarray, np.ndarray, EllipseFeatureBundle]]:
bgr = cv2.imread(str(image_path), cv2.IMREAD_COLOR)
if bgr is None:
raise FileNotFoundError(str(image_path))
return extract_ellipse_keypoints_from_bgr(bgr, **kwargs)
class ResNetFeatureExtractor(nn.Module):
"""基于 ResNet18 多层特征图的双线性插值采样器。"""
def __init__(
self,
checkpoint_path: Optional[PathLike] = None,
device: Optional[torch.device] = None,
):
super().__init__()
self.device = device or _default_device()
resnet = models.resnet18(weights=None)
ckpt = Path(checkpoint_path) if checkpoint_path else DEFAULT_CHECKPOINT
if ckpt.is_file():
state = torch.load(ckpt, map_location=self.device, weights_only=False)
state_dict = state.get("model_state_dict", state)
cleaned = {}
for k, v in state_dict.items():
name = k[7:] if k.startswith("module.") else k
cleaned[name] = v
resnet.load_state_dict(cleaned, strict=False)
print(f"[ok] loaded checkpoint: {ckpt}")
else:
print(f"[warn] checkpoint not found: {ckpt}, using random init")
self.conv1 = resnet.conv1
self.bn1 = resnet.bn1
self.relu = resnet.relu
self.maxpool = resnet.maxpool
self.layer1 = resnet.layer1
self.layer2 = resnet.layer2
self.layer3 = resnet.layer3
self.layer4 = resnet.layer4
self.feature_layers = nn.ModuleDict({
"layer1": self.layer1,
"layer2": self.layer2,
"layer3": self.layer3,
})
self.strides = {"layer1": 4, "layer2": 8, "layer3": 16}
self.out_channels = {"layer1": 64, "layer2": 128, "layer3": 256}
self.total_dim = sum(self.out_channels.values())
def forward(self, images: torch.Tensor, keypoints: torch.Tensor) -> torch.Tensor:
"""
Parameters
----------
images : [B, 3, H, W]
keypoints : [B, N, 2],像素坐标 (x, y)
Returns
-------
[B, N, total_dim]
"""
x = self.conv1(images)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
feature_maps = {}
for name, layer in self.feature_layers.items():
x = layer(x)
feature_maps[name] = x
sampled_features_list = []
for name, fmap in feature_maps.items():
stride = self.strides[name]
_, _, fh, fw = fmap.shape
scaled_kps = keypoints / stride
if fw > 1:
norm_x = (scaled_kps[..., 0] / (fw - 1)) * 2 - 1
else:
norm_x = torch.zeros_like(scaled_kps[..., 0])
if fh > 1:
norm_y = (scaled_kps[..., 1] / (fh - 1)) * 2 - 1
else:
norm_y = torch.zeros_like(scaled_kps[..., 1])
norm_x = torch.clamp(norm_x, -1, 1)
norm_y = torch.clamp(norm_y, -1, 1)
grid = torch.stack([norm_x, norm_y], dim=-1).unsqueeze(2)
sampled = F.grid_sample(
fmap, grid, mode="bilinear", align_corners=True, padding_mode="border"
)
sampled = sampled.squeeze(-1).permute(0, 2, 1)
sampled_features_list.append(sampled)
return torch.cat(sampled_features_list, dim=-1)
@torch.no_grad()
def extract_resnet_point_features(
rgb_bgr: np.ndarray,
points_xy: np.ndarray,
extractor: Optional[ResNetFeatureExtractor] = None,
checkpoint_path: Optional[PathLike] = None,
device: Optional[torch.device] = None,
) -> torch.Tensor:
"""单张图多点采样,返回 (N, RESNET_DEEP_RAW_DIM)。"""
device = device or _default_device()
if extractor is None:
extractor = ResNetFeatureExtractor(checkpoint_path=checkpoint_path, device=device)
extractor.to(device)
extractor.eval()
image_tensor, _, _ = _rgb_bgr_to_tensor(rgb_bgr, device)
kps = torch.from_numpy(points_xy.astype(np.float32)).unsqueeze(0).to(device)
_, _, h, w = image_tensor.shape
kps[..., 0] = kps[..., 0].clamp(0, w - 1)
kps[..., 1] = kps[..., 1].clamp(0, h - 1)
return extractor(image_tensor, kps).squeeze(0)
@torch.no_grad()
def extract_resnet_point_features_batch(
rgb_bgr_list: Sequence[np.ndarray],
points_xy_list: Sequence[np.ndarray],
extractor: Optional[ResNetFeatureExtractor] = None,
checkpoint_path: Optional[PathLike] = None,
device: Optional[torch.device] = None,
) -> torch.Tensor:
"""
批量多图多点采样。
Parameters
----------
rgb_bgr_list : 长度 B,每张图 resize 后尺寸需一致
points_xy_list : 长度 B,每项 (N, 2);通常 N=360
Returns
-------
(B, N, RESNET_DEEP_RAW_DIM)
"""
if len(rgb_bgr_list) != len(points_xy_list):
raise ValueError("rgb_bgr_list 与 points_xy_list 长度须一致")
if not rgb_bgr_list:
raise ValueError("输入列表为空")
device = device or _default_device()
if extractor is None:
extractor = ResNetFeatureExtractor(checkpoint_path=checkpoint_path, device=device)
extractor.to(device)
extractor.eval()
images = _batch_rgb_bgr_to_tensor(rgb_bgr_list, device)
_, _, h, w = images.shape
kps = torch.stack([
torch.from_numpy(pts.astype(np.float32)) for pts in points_xy_list
], dim=0).to(device)
kps[..., 0] = kps[..., 0].clamp(0, w - 1)
kps[..., 1] = kps[..., 1].clamp(0, h - 1)
return extractor(images, kps)
def project_deep_features(
raw_features: torch.Tensor,
proj: nn.Linear,
) -> torch.Tensor:
"""§4.1.3 线性投影:(*, RESNET_DEEP_RAW_DIM) → (*, d_deep)。"""
return proj(raw_features)
def draw_keypoints_on_bgr(
bgr: np.ndarray,
keypoints_xy: np.ndarray,
interp_flag: Optional[np.ndarray] = None,
step: int = 1,
point_radius: int = 2,
draw_index: bool = False,
draw_trajectory: bool = True,
bundle: Optional[EllipseFeatureBundle] = None,
) -> np.ndarray:
"""
在 BGR 图像上绘制 kps_list 中的采样点。
Parameters
----------
bgr : 与 keypoints_xy 同一坐标系的图像(通常为 resize 后的 rgb)
keypoints_xy : (N, 2),(x, y) 像素坐标
interp_flag : (N,) 可选,1=插值点,0=实测点;None 则全部同色
"""
vis = bgr.copy()
h, w = vis.shape[:2]
n = keypoints_xy.shape[0]
if bundle is not None:
box = cv2.boxPoints(bundle.ellipse_cv)
cv2.polylines(vis, [np.int32(box)], True, (0, 200, 0), 1, cv2.LINE_AA)
if draw_trajectory and n > 1:
pts = []
for t in range(0, n, max(1, step)):
x, y = int(round(keypoints_xy[t, 0])), int(round(keypoints_xy[t, 1]))
if 0 <= x < w and 0 <= y < h:
pts.append([x, y])
if len(pts) > 1:
cv2.polylines(vis, [np.int32(pts)], False, (255, 180, 0), 1, cv2.LINE_AA)
for t in range(0, n, max(1, step)):
x, y = int(round(keypoints_xy[t, 0])), int(round(keypoints_xy[t, 1]))
if not (0 <= x < w and 0 <= y < h):
continue
if interp_flag is not None and interp_flag[t] == 1:
color = (0, 165, 255) # 插值:橙
else:
color = (0, 0, 255) # 实测:红
cv2.circle(vis, (x, y), point_radius, color, -1, cv2.LINE_AA)
if t == 0:
cv2.circle(vis, (x, y), point_radius + 2, (0, 255, 0), 1, cv2.LINE_AA)
if draw_index and t % max(1, step * 4) == 0:
cv2.putText(
vis, str(t), (x + 2, y - 2),
cv2.FONT_HERSHEY_SIMPLEX, 0.25, (255, 255, 255), 1, cv2.LINE_AA,
)
return vis
def scale_keypoints_to_image(
keypoints_xy: np.ndarray,
src_shape: Tuple[int, int],
dst_shape: Tuple[int, int],
) -> np.ndarray:
"""将关键点从 src 图像坐标系缩放到 dst 图像坐标系。shape 为 (H, W)。"""
src_h, src_w = src_shape
dst_h, dst_w = dst_shape
scaled = keypoints_xy.astype(np.float32).copy()
scaled[:, 0] *= dst_w / src_w
scaled[:, 1] *= dst_h / src_h
return scaled
def visualize_kps_list(
rgb_list: Sequence[np.ndarray],
kps_list: Sequence[np.ndarray],
out_dir: Union[str, Path],
orig_bgr_list: Optional[Sequence[np.ndarray]] = None,
bundles: Optional[Sequence[EllipseFeatureBundle]] = None,
names: Optional[Sequence[str]] = None,
step: int = 2,
) -> List[Path]:
"""
将 kps_list 画到对应图像上并保存。
每张图输出两版:
- *_kps_proc.png : resize 后输入图(与 ResNet 采样坐标一致)
- *_kps_orig.png : 原始分辨率图(关键点已按比例缩放)
"""
out_root = Path(out_dir)
out_root.mkdir(parents=True, exist_ok=True)
saved: List[Path] = []
for i, (rgb, kps) in enumerate(zip(rgb_list, kps_list)):
name = names[i] if names else f"sample_{i:03d}"
stem = Path(name).stem if "/" in name or "\\" in name else name
interp = None
bundle = bundles[i] if bundles is not None else None
if bundle is not None:
interp = bundle.sequence.interp_flag
proc_vis = draw_keypoints_on_bgr(
rgb, kps, interp_flag=interp, step=step, bundle=bundle,
)
proc_path = out_root / f"{stem}_kps_proc.png"
cv2.imwrite(str(proc_path), proc_vis)
saved.append(proc_path)
if orig_bgr_list is not None:
orig = orig_bgr_list[i]
kps_orig = scale_keypoints_to_image(
kps, rgb.shape[:2], orig.shape[:2],
)
orig_vis = draw_keypoints_on_bgr(
orig, kps_orig, interp_flag=interp, step=step,
point_radius=max(2, int(round(min(orig.shape[:2]) / 80))),
bundle=None,
)
orig_path = out_root / f"{stem}_kps_orig.png"
cv2.imwrite(str(orig_path), orig_vis)
saved.append(orig_path)
return saved
def run_batch_test(
image_dir: Union[str, Path] = "data/images",
n_samples: int = 3,
seed: int = 42,
img_size: int = DEFAULT_IMG_SIZE,
checkpoint_path: Optional[PathLike] = None,
out_dir: Union[str, Path] = "outputs/resnet_point_feat_test",
) -> None:
"""从目录随机抽多张图,用椭圆特征关键点批量提取 ResNet 深度特征。"""
root = Path(image_dir)
paths = sorted(root.glob("*.png")) + sorted(root.glob("*.jpg"))
if not paths:
raise FileNotFoundError(f"目录下无图像: {image_dir}")
random.seed(seed)
picks = random.sample(paths, min(n_samples, len(paths)))
device = _default_device()
print(f"[device] {device}")
print(f"[images] ({len(picks)}): {[p.name for p in picks]}")
rgb_list: List[np.ndarray] = []
kps_list: List[np.ndarray] = []
orig_bgr_list: List[np.ndarray] = []
bundle_list: List[EllipseFeatureBundle] = []
meta: List[dict] = []
for p in picks:
orig_bgr = cv2.imread(str(p), cv2.IMREAD_COLOR)
if orig_bgr is None:
print(f" [skip] {p.name}: cannot read image")
continue
result = extract_ellipse_keypoints_from_bgr(orig_bgr, img_size=img_size)
if result is None:
print(f" [skip] {p.name}: valid ellipse points insufficient")
continue
rgb, kps, bundle = result
valid = int(np.sum(bundle.sequence.interp_flag == 0))
rgb_list.append(rgb)
kps_list.append(kps)
orig_bgr_list.append(orig_bgr)
bundle_list.append(bundle)
meta.append({"path": str(p), "valid_points": valid, "n_theta": kps.shape[0]})
print(f" [ok] {p.name}: {kps.shape[0]} pts, measured {valid}")
if not rgb_list:
raise RuntimeError("无可用样本,批量测试失败")
out_root = Path(out_dir)
out_root.mkdir(parents=True, exist_ok=True)
viz_paths = visualize_kps_list(
rgb_list,
kps_list,
out_dir=out_root / "kps_vis",
orig_bgr_list=orig_bgr_list,
bundles=bundle_list,
names=[Path(m["path"]).name for m in meta],
step=2,
)
print(f"[viz] saved {len(viz_paths)} keypoint overlay images -> {out_root / 'kps_vis'}")
for vp in viz_paths:
print(f" {vp.name}")
extractor = ResNetFeatureExtractor(checkpoint_path=checkpoint_path, device=device)
extractor.to(device)
extractor.eval()
features = extract_resnet_point_features_batch(
rgb_list, kps_list, extractor=extractor, device=device
)
print("\n[batch result]")
print(f" batch size: {len(rgb_list)}")
print(f" keypoints (logical): [{features.shape[0]}, {features.shape[1]}, 2]")
print(f" features: {tuple(features.shape)} -> [B, N, {RESNET_DEEP_RAW_DIM}]")
print(f" expected dim: {extractor.total_dim}")
if features.shape[-1] == extractor.total_dim:
print("[ok] dimension match")
else:
print("[fail] dimension mismatch")
for i, m in enumerate(meta):
fi = features[i]
print(
f" [{i}] {Path(m['path']).name}: "
f"feat mean={fi.mean().item():.4f}, std={fi.std().item():.4f}"
)
out_root = Path(out_dir)
out_root.mkdir(parents=True, exist_ok=True)
np.save(out_root / "batch_deep_features.npy", features.cpu().numpy())
np.save(out_root / "batch_keypoints.npy", np.stack(kps_list, axis=0))
with open(out_root / "batch_meta.txt", "w", encoding="utf-8") as f:
for i, m in enumerate(meta):
f.write(f"{i}\t{m['path']}\tvalid={m['valid_points']}\n")
print(f"[saved] {out_root}")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="ResNet 多点深度特征批量测试")
parser.add_argument("--image-dir", default="data/images")
parser.add_argument("--n", type=int, default=30, help="随机测试图像数量")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--img-size", type=int, default=DEFAULT_IMG_SIZE)
parser.add_argument("--checkpoint", default=None)
parser.add_argument("--out-dir", default="outputs/resnet_point_feat_test")
args = parser.parse_args()
run_batch_test(
image_dir=args.image_dir,
n_samples=args.n,
seed=args.seed,
img_size=args.img_size,
checkpoint_path=args.checkpoint,
out_dir=args.out_dir,
)