滑动窗口提取图像保存至3D矩阵时报错，求排查解决

Let's walk through the issues causing errors in your code and how to resolve them:

1. Primary Issue: Index Out-of-Bounds #

You initialized matrix A with a fixed size of (100, winW, winH), but the number of sliding windows generated by your pyramid and sliding window loops will almost certainly exceed 100. When the variable b increments to 101, trying to assign to A[b,:,:] will throw an index error—since Python uses 0-based indexing, A only accepts indices from 0 to 99 for its first dimension.

Additionally, starting with b=1 wastes the first position (A[0,:,:]) and makes you hit the index limit even faster.

2. Secondary Issue: Potential Data Type Mismatch #

np.empty() creates an array with a default float64 data type, but your window is a grayscale image slice stored as uint8 (integer values 0-255). Directly assigning uint8 values to a float64 array can lead to unnecessary type conversions or unexpected behavior.

Recommended Solution: Use a Dynamic List (Flexible & Simple) #

Instead of preallocating a fixed-size matrix (which requires knowing the exact number of windows upfront), use a Python list to dynamically collect all valid windows, then convert it to a 3D numpy array once you're done. This avoids index errors entirely and handles any number of windows.

Modified Code Snippets: #

Replace your matrix initialization section:

# Replace fixed-size matrix with a dynamic list
windows_list = []

Then, in the sliding window loop, replace the problematic lines with:

# Add the valid window to our list
windows_list.append(window)

After processing all windows, convert the list to a 3D numpy array:

# Convert the list to a 3D matrix (shape: [number_of_windows, winH, winW])
A = np.array(windows_list)
# Verify the shape (optional)
print(f"Final shape of matrix A: {A.shape}")

Alternative: Preallocate a Fixed-Size Matrix (If You Need It) #

If you must use a preallocated matrix, first calculate the exact number of valid windows to set the correct size:

# First, count all valid windows
window_count = 0
for resized in pyramid(image, scale=2):
    for (x, y, window) in sliding_window(resized, stepSize=32, windowSize=(winW, winH)):
        if window.shape[0] != winH or window.shape[1] != winW:
            continue
        window_count += 1

# Initialize matrix with matching size and data type
A = np.empty((window_count, winH, winW), dtype=window.dtype)
b = 0  # Start at 0 for 0-based indexing

Then in the loop:

A[b,:,:] = window
b += 1

This ensures you never exceed the matrix bounds and matches the data type of your windows.

内容的提问来源于stack exchange，提问作者Aros