deepellipse-circleid / code /circle_feature.py
qinxunhui's picture
Add small artifacts (code, outputs, configs, LICENSE)
519ab98 verified
Raw History Blame Contribute Delete
10.3 kB
import cv2
import numpy
import numpy as np
import torch
import resnet_point_feature
from old import sift_feature, polar_order
backbone_device = "cpu" # device used by the augmentation-side feature path
backbone = None # built lazily by get_backbone()
def get_backbone():
"""Lazily build the ResNet point-feature extractor.
Built on first use so that merely importing this module -- which the
released inference path does transitively -- does not attempt to load a
backbone checkpoint at import time.
"""
global backbone
if backbone is None:
backbone = resnet_point_feature.ResNetFeatureExtractor().to(backbone_device)
backbone.eval()
return backbone
class Config:
TARGET_SIZE = 100
SIFT_EXTRACTOR = sift_feature.SIFTFeatureExtractorOptimized()
SUCCESS_POINT_RATION = 0.2
def exact_features(image):
features_dict = polar_order.process_single_image_polar(image)
polar_order_features = features_dict["features"] # (120, 8)
ellipse_feature = numpy.array([features_dict["ellipse_params"]['cx'] / Config.TARGET_SIZE,
features_dict["ellipse_params"]['cy'] / Config.TARGET_SIZE,
features_dict["ellipse_params"]['MA'] / Config.TARGET_SIZE,
features_dict["ellipse_params"]['ma'] / Config.TARGET_SIZE,
features_dict["ellipse_params"]['angle'] / 360.0,
features_dict["ellipse_params"]['fit_error'] / Config.TARGET_SIZE,
])
# sift_descriptors np.ndarray 120*32,
# success_mask, np.ndarray,120
sift_descriptors, success_mask = sift_feature.extract_sift_for_single_image_pure(image, features_dict[
"matched_points"], Config.SIFT_EXTRACTOR)
success_point_ratio = sum(success_mask == True) / len(success_mask)
if success_point_ratio < Config.SUCCESS_POINT_RATION:
print(success_point_ratio)
raise ValueError("边缘点过少")
xy_feature = features_dict["matched_points"] / Config.TARGET_SIZE
ellipse_feature_expanded = np.tile(ellipse_feature, (120, 1))
# 2. 在第一维度(列方向)上拼接所有特征
all_features = np.concatenate([
polar_order_features, # (120, 8)
ellipse_feature_expanded, # (120, 6)
xy_feature, # (120, 2)
sift_descriptors # (120, 32)
], axis=1)
return all_features
def extract_features_diff(image, bgr_norm_image=None):
"""
提取增强版融合特征 (包含差分特征)
修复:处理角度特征的周期性跳变问题
"""
try:
# ================= 1. 基础特征提取 =================
# 1.1 极坐标特征 (120, 8)
# 假设格式: [t_r, t_theta, r, theta, e_r, e_theta, dis, strategy]
# 其中 t_theta (idx=1) 和 theta (idx=3) 是归一化角度 (0~1)
features_dict = polar_order.process_single_image_polar(image)
if features_dict is None:
return None
polar_order_features = features_dict["features"].astype(np.float32) # (120, 8)
# 1.2 椭圆全局特征 (6,) -> 归一化
ellipse_params = features_dict["ellipse_params"]
ellipse_feature = np.array([
ellipse_params['cx'] / Config.TARGET_SIZE,
ellipse_params['cy'] / Config.TARGET_SIZE,
ellipse_params['MA'] / Config.TARGET_SIZE,
ellipse_params['ma'] / Config.TARGET_SIZE,
ellipse_params['angle'] / 360.0,
ellipse_params['fit_error'] / Config.TARGET_SIZE,
], dtype=np.float32)
if bgr_norm_image is None:
sift_descriptors, success_mask = sift_feature.extract_sift_for_single_image_pure(
image, features_dict["matched_points"], Config.SIFT_EXTRACTOR
)
else:
coords = []
mask = []
for pt in features_dict["matched_points"]:
if pt is not None:
coords.append([float(pt[0]), float(pt[1])])
mask.append(True)
else:
coords.append([0, 0])
mask.append(False)
success_mask = np.array(mask, dtype=bool)
coords_np = np.array(coords, dtype=np.float32)
keypoints_tensor = torch.from_numpy(coords_np).float().to(backbone_device)
# 再次检查范围,确保没有超出新图像边界
keypoints_tensor[..., 0] = torch.clamp(keypoints_tensor[..., 0], 0, image.shape[0] - 1)
keypoints_tensor[..., 1] = torch.clamp(keypoints_tensor[..., 1], 0, image.shape[1] - 1)
bgr_norm_image_tensor = torch.from_numpy(bgr_norm_image.astype(np.float32)).unsqueeze(0).permute(0, 3, 1, 2).to(backbone_device)
keypoints_tensor = keypoints_tensor.unsqueeze(0)
dense_point_feature = get_backbone()(bgr_norm_image_tensor, keypoints_tensor) #B *N *D
# 适配下列代码,改名改格式
mask_expanded = torch.from_numpy(success_mask).to(backbone_device).unsqueeze(0).unsqueeze(-1)
filtered_features = dense_point_feature * mask_expanded
sift_descriptors = filtered_features.squeeze(0).detach().cpu().numpy()
success_point_ratio = np.sum(success_mask) / len(success_mask)
if success_point_ratio < Config.SUCCESS_POINT_RATION:
# raise ValueError(f"边缘点过少:{success_point_ratio:.2f} < {Config.SUCCESS_POINT_RATION}")
return None
# 1.4 坐标特征 (120, 2) -> 归一化
matched_points = features_dict["matched_points"]
xy_feature = matched_points / Config.TARGET_SIZE # (120, 2)
# 1.5 扩展椭圆特征 (120, 6)
ellipse_feature_expanded = np.tile(ellipse_feature, (120, 1))
# ================= 2. 差分特征计算 (Delta Features) =================
def compute_circular_diff(features_2d, is_angle=False):
"""
计算循环差分,支持角度周期性修正
Args:
features_2d: (N, D) 输入特征
is_angle: 是否为角度特征 (0~1 周期)
Returns:
delta: (N, D) 差分值
"""
# 获取前一个时刻的特征 (循环移位)
prev_features = np.roll(features_2d, shift=1, axis=0)
# 原始差分
delta = features_2d - prev_features
if is_angle:
# 🔧 角度周期性修正
# 原理:角度差值范围应在 [-0.5, 0.5] 之间 (即 -180° ~ 180°)
# 如果 diff > 0.5 (如 0.9), 说明跨过了 0°/360° 边界,实际应为 diff - 1 (如 -0.1)
# 如果 diff < -0.5 (如 -0.9), 说明跨过了边界,实际应为 diff + 1 (如 0.1)
delta = (delta + 0.5) % 1.0 - 0.5
return delta.astype(np.float32)
# 2.1 Polar 差分 (前 6 维)
# 索引说明: 0=t_r, 1=t_theta(角度), 2=r, 3=theta(角度), 4=e_r, 5=e_theta
polar_geo_features = polar_order_features[:, :6] # (120, 6)
# 分离角度列和非角度列
polar_non_angle = polar_geo_features[:, [0, 2, 4]] # (120, 3) -> r 相关
polar_angle = polar_geo_features[:, [1, 3, 5]] # (120, 3) -> theta 相关
# 分别计算差分
polar_non_angle_delta = compute_circular_diff(polar_non_angle, is_angle=False)
polar_angle_delta = compute_circular_diff(polar_angle, is_angle=True)
# 恢复原始列顺序 [0, 1, 2, 3, 4, 5]
polar_delta = np.zeros_like(polar_geo_features)
polar_delta[:, [0, 2, 4]] = polar_non_angle_delta
polar_delta[:, [1, 3, 5]] = polar_angle_delta
# 2.2 XY 坐标差分 (2 维)
# 坐标是笛卡尔空间,不需要周期性修正 (除非图像是环面的,但这里不是)
xy_delta = compute_circular_diff(xy_feature, is_angle=False) # (120, 2)
# 2.3 SIFT 描述子差分 (32 维)
# SIFT 是向量空间,不需要周期性修正
sift_delta = compute_circular_diff(sift_descriptors, is_angle=False) # (120, 32)
# ================= 3. 特征拼接 =================
# 3.1 静态特征拼接 (48 维)
static_features = np.concatenate([
polar_order_features, # (120, 8)
ellipse_feature_expanded, # (120, 6)
xy_feature, # (120, 2)
sift_descriptors # (120, 32) or (120, 448)
], axis=1)
# 3.2 动态差分特征拼接 (40 维)
dynamic_features = np.concatenate([
polar_delta, # (120, 6) - 已修复角度跳变
xy_delta, # (120, 2)
# sift_delta # (120, 32) or (120, 448)
], axis=1)
# 3.3 最终融合 (88 维)
all_features = np.concatenate([
static_features,
dynamic_features
], axis=1)
# 验证维度
# assert all_features.shape == (120, 88), f"特征维度错误:{all_features.shape}"
assert all_features.shape == (120, 472), f"特征维度错误:{all_features.shape}"
# 🔍 可选:调试检查角度差分是否合理
# max_angle_diff = np.max(np.abs(polar_delta[:, [1, 3, 5]]))
# assert max_angle_diff <= 0.5, f"角度差分异常,最大值超过 0.5 (180 度): {max_angle_diff}"
return all_features
except Exception as e:
print(f"特征提取失败:{str(e)}")
import traceback
traceback.print_exc()
return None
if __name__ == "__main__":
image_path = "/path/to/circle_id/images/images/00059.png"
img = cv2.imread(image_path)
# fea = exact_features(img)
# print(fea.shape)
image_resized = cv2.resize(img, (100, 100), interpolation=cv2.INTER_AREA)
NORM_MEAN_IMG = np.array([0.485, 0.456, 0.406], dtype=np.float32).reshape(1, 1, -1)
NORM_STD_IMG = np.array([0.229, 0.224, 0.225], dtype=np.float32).reshape(1, 1, -1)
# --- 处理 Image ---
image = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img_norm = (image / 255.0 - NORM_MEAN_IMG) / NORM_STD_IMG
fea = extract_features_diff(img, img_norm.astype(np.float32))