如何通过OpenCV轮廓从UE4带Mask图像中裁剪目标并获取位置

Hey there! Let's fix that inconsistent contour detection issue you're facing. Since you already have dedicated mask images (not trying to segment directly from RGB), we can leverage that to build a much more robust workflow—no more guessing with threshold values or dealing with missing/wrong contours.

First: Let's Use the Mask's Strength Directly #

UnrealCV masks are designed to clearly separate your target from the background, so we don't need to mess around with generic thresholding. The simplest and most reliable way is to just find all non-zero pixels in the mask (since your target should be the only non-background region):

import numpy as np
import cv2
import os

# Load your images
rgb = cv2.imread('image_path/rgb_1.png')
mask = cv2.imread('image_path/mask_1.png')

# Convert mask to grayscale (even if it's single-channel, this ensures consistency)
gray_mask = cv2.cvtColor(mask, cv2.COLOR_BGR2GRAY)

# Find all non-zero pixels in the mask
non_zero_coords = np.nonzero(gray_mask)

if len(non_zero_coords[0]) > 0:
    # Calculate the tight bounding box around the target
    y_min, y_max = np.min(non_zero_coords[0]), np.max(non_zero_coords[0])
    x_min, x_max = np.min(non_zero_coords[1]), np.max(non_zero_coords[1])
    
    # Draw the box for visualization
    boxed_mask = cv2.rectangle(mask.copy(), (x_min, y_min), (x_max, y_max), (255,255,255), 1)
    boxed_rgb = cv2.rectangle(rgb.copy(), (x_min, y_min), (x_max, y_max), (255,255,255), 1)
    
    # Crop the target from RGB
    cropped_target = rgb[y_min:y_max+1, x_min:x_max+1]
    
    # Show results
    cv2.imshow('Mask with Bounding Box', boxed_mask)
    cv2.imshow('RGB with Bounding Box', boxed_rgb)
    cv2.imshow('Cropped Target', cropped_target)
    
    if cv2.waitKey(0) == ord('q'):
        cv2.destroyAllWindows()
else:
    print("No target pixels detected in the mask!")

cv2.destroyAllWindows()

Why this works better: #

• No manual threshold values to tweak—it directly uses the mask's intended purpose.
• Avoids issues with contour detection missing the target or picking up noise.

Fixing Noisy Masks #

If your mask has random small noise spots (which can mess up the non-zero pixel check), add a quick morphological operation to clean it up first:

# Add this right after converting to grayscale
kernel = np.ones((3, 3), np.uint8)
# "Opening" operation: erode then dilate to remove small noise
clean_mask = cv2.morphologyEx(gray_mask, cv2.MORPH_OPEN, kernel)

# Now use clean_mask instead of gray_mask to find non-zero pixels
non_zero_coords = np.nonzero(clean_mask)

If You Still Want to Use Contour Detection #

If you prefer contours (maybe you need more shape info later), let's fix that workflow too:

• Use Otsu's thresholding to auto-pick the best threshold value (no more 127 guesswork).
• Only detect external contours (ignore nested ones).
• Sort contours by area and pick the largest one (your target is definitely the biggest shape).

import numpy as np
import cv2

rgb = cv2.imread('image_path/rgb_1.png')
mask = cv2.imread('image_path/mask_1.png')
gray_mask = cv2.cvtColor(mask, cv2.COLOR_BGR2GRAY)

# Otsu's thresholding automatically finds the optimal threshold
ret, thresh = cv2.threshold(gray_mask, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)

# Find only external contours (no nested ones)
contours, hierarchy = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

if contours:
    # Sort contours by area (descending) and pick the largest one
    sorted_contours = sorted(contours, key=cv2.contourArea, reverse=True)
    target_contour = sorted_contours[0]
    
    # Get bounding box
    x, y, w, h = cv2.boundingRect(target_contour)
    
    # Draw and crop (same as before)
    boxed_mask = cv2.rectangle(mask.copy(), (x, y), (x+w, y+h), (255,255,255), 1)
    boxed_rgb = cv2.rectangle(rgb.copy(), (x, y), (x+w, y+h), (255,255,255), 1)
    cropped_target = rgb[y:y+h, x:x+w]
    
    cv2.imshow('Contour Box on Mask', boxed_mask)
    cv2.imshow('Cropped Target', cropped_target)
    if cv2.waitKey(0) == ord('q'):
        cv2.destroyAllWindows()
else:
    print("No contours detected!")

cv2.destroyAllWindows()

Quick Check for Your Mask #

Double-check what your mask's pixel values are:

• If your target is black and background is white, change np.nonzero(gray_mask) to np.nonzero(255 - gray_mask) (inverts the mask).
• Ensure your mask isn't compressed with lossy formats (like JPEG) which can introduce artifacts—save masks as PNG if possible.

These methods should work for all your single-target masks, since they rely on the mask's core purpose of isolating the target.

内容的提问来源于stack exchange，提问作者karim