DINOv3 Similarity Visualization¶

This notebook demonstrates the new visualization functionality for DINOv3 similarity analysis in the geoai package.

Install packages¶

To use the new functionality, ensure the required packages are installed.

In [ ]:

# %pip install geoai-py

# %pip install geoai-py

Import libraries¶

In [ ]:

import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
import geoai
from geoai.dinov3 import DINOv3GeoProcessor, visualize_similarity_results

%matplotlib inline

import numpy as np import matplotlib.pyplot as plt from PIL import Image import geoai from geoai.dinov3 import DINOv3GeoProcessor, visualize_similarity_results %matplotlib inline

In [ ]:

url = "https://huggingface.co/spaces/pszemraj/dinov3-viz-sat493m/resolve/main/aralsea_tmo_2000238_lrg.jpg"
test_image_path = geoai.download_file(url)

url = "https://huggingface.co/spaces/pszemraj/dinov3-viz-sat493m/resolve/main/aralsea_tmo_2000238_lrg.jpg" test_image_path = geoai.download_file(url)

In [ ]:

# Create and display test image
test_image = Image.open(test_image_path)

plt.figure(figsize=(8, 8))
plt.imshow(test_image)
plt.axis("off")
plt.show()

# Create and display test image test_image = Image.open(test_image_path) plt.figure(figsize=(8, 8)) plt.imshow(test_image) plt.axis("off") plt.show()

Method 1: Quick Visualization with visualize_similarity_results()¶

The easiest way to create and visualize similarity results in one step.

In [ ]:

# Query coordinates
query_coords = (1250, 1250)

# Side-by-side visualization
results = visualize_similarity_results(
    input_image=test_image_path,
    query_coords=query_coords,
    overlay=False,  # Side-by-side view
    figsize=(15, 6),
    colormap="turbo",
    show_query_point=True,
)

plt.show()
print(f"Query patch coordinates: {results['patch_coords']}")
print(f"Patch grid size: {results['patch_grid_size']}")

# Query coordinates query_coords = (1250, 1250) # Side-by-side visualization results = visualize_similarity_results( input_image=test_image_path, query_coords=query_coords, overlay=False, # Side-by-side view figsize=(15, 6), colormap="turbo", show_query_point=True, ) plt.show() print(f"Query patch coordinates: {results['patch_coords']}") print(f"Patch grid size: {results['patch_grid_size']}")

In [ ]:

# Overlay visualization
results_overlay = visualize_similarity_results(
    input_image=test_image_path,
    query_coords=query_coords,
    overlay=True,  # Overlay view
    alpha=0.6,  # Transparency
    colormap="jet",
    show_query_point=True,
)

plt.show()

# Overlay visualization results_overlay = visualize_similarity_results( input_image=test_image_path, query_coords=query_coords, overlay=True, # Overlay view alpha=0.6, # Transparency colormap="jet", show_query_point=True, ) plt.show()

Method 2: Step-by-step with DINOv3GeoProcessor¶

For more control over the process, you can use the processor methods directly.

In [ ]:

# Initialize processor
processor = DINOv3GeoProcessor()

# Load and preprocess image
data, metadata = processor.load_image(test_image_path)
image = processor.preprocess_image_for_dinov3(data)
features, h_patches, w_patches = processor.extract_features(image)

print(f"Image size: {image.size}")
print(f"Features shape: {features.shape}")
print(f"Patch grid: {h_patches} x {w_patches}")

# Initialize processor processor = DINOv3GeoProcessor() # Load and preprocess image data, metadata = processor.load_image(test_image_path) image = processor.preprocess_image_for_dinov3(data) features, h_patches, w_patches = processor.extract_features(image) print(f"Image size: {image.size}") print(f"Features shape: {features.shape}") print(f"Patch grid: {h_patches} x {w_patches}")

In [ ]:

# Convert pixel coordinates to patch coordinates
img_w, img_h = data.shape[2], data.shape[1]
patch_x = int((query_coords[0] / img_w) * w_patches)
patch_y = int((query_coords[1] / img_h) * h_patches)

print(f"Query pixel coordinates: {query_coords}")
print(f"Query patch coordinates: ({patch_x}, {patch_y})")

# Compute similarity
similarities = processor.compute_patch_similarity(features, patch_x, patch_y)
similarity_array = similarities.cpu().numpy()

print(f"Similarity range: {similarity_array.min():.3f} - {similarity_array.max():.3f}")

# Convert pixel coordinates to patch coordinates img_w, img_h = data.shape[2], data.shape[1] patch_x = int((query_coords[0] / img_w) * w_patches) patch_y = int((query_coords[1] / img_h) * h_patches) print(f"Query pixel coordinates: {query_coords}") print(f"Query patch coordinates: ({patch_x}, {patch_y})") # Compute similarity similarities = processor.compute_patch_similarity(features, patch_x, patch_y) similarity_array = similarities.cpu().numpy() print(f"Similarity range: {similarity_array.min():.3f} - {similarity_array.max():.3f}")

In [ ]:

# Visualize patch grid
fig = processor.visualize_patches(
    image=image, features=features, patch_coords=(patch_x, patch_y), figsize=(12, 8)
)

plt.show()

# Visualize patch grid fig = processor.visualize_patches( image=image, features=features, patch_coords=(patch_x, patch_y), figsize=(12, 8) ) plt.show()

In [ ]:

# Create side-by-side visualization
fig = processor.visualize_similarity(
    source=test_image_path,
    similarity_data=similarity_array,
    query_coords=query_coords,
    patch_coords=(patch_x, patch_y),
    overlay=False,
    colormap="jet",
    figsize=(15, 6),
)

plt.show()

# Create side-by-side visualization fig = processor.visualize_similarity( source=test_image_path, similarity_data=similarity_array, query_coords=query_coords, patch_coords=(patch_x, patch_y), overlay=False, colormap="jet", figsize=(15, 6), ) plt.show()

In [ ]:

# Create overlay visualization
fig = processor.visualize_similarity(
    source=test_image_path,
    similarity_data=similarity_array,
    query_coords=query_coords,
    overlay=True,
    alpha=0.5,
    colormap="jet",
)

plt.show()

# Create overlay visualization fig = processor.visualize_similarity( source=test_image_path, similarity_data=similarity_array, query_coords=query_coords, overlay=True, alpha=0.5, colormap="jet", ) plt.show()

Method 3: Create Overlay Image Array¶

You can also create overlay images programmatically for further processing.

In [ ]:

# Create overlay as numpy array
overlay_img = processor.create_similarity_overlay(
    source=test_image_path, similarity_data=similarity_array, colormap="jet", alpha=0.6
)

# Display the overlay
plt.figure(figsize=(10, 10))
plt.imshow(overlay_img)
plt.title("Programmatically Created Overlay")
plt.axis("off")
plt.show()

# Save overlay as image file
overlay_pil = Image.fromarray((overlay_img * 255).astype(np.uint8))
overlay_pil.save("demo_overlay.png")
print("Overlay saved to demo_overlay.png")

# Create overlay as numpy array overlay_img = processor.create_similarity_overlay( source=test_image_path, similarity_data=similarity_array, colormap="jet", alpha=0.6 ) # Display the overlay plt.figure(figsize=(10, 10)) plt.imshow(overlay_img) plt.title("Programmatically Created Overlay") plt.axis("off") plt.show() # Save overlay as image file overlay_pil = Image.fromarray((overlay_img * 255).astype(np.uint8)) overlay_pil.save("demo_overlay.png") print("Overlay saved to demo_overlay.png")

Test Different Query Points¶

Let's test similarity analysis with different query points.

In [ ]:

# Test multiple query points
query_points = [
    (128, 128),  # Red circle
    (384, 128),  # Green circle
    (128, 384),  # Blue circle
    (384, 384),  # Yellow circle
]

colors = ["Red", "Green", "Blue", "Yellow"]

fig, axes = plt.subplots(2, 2, figsize=(20, 16))
axes = axes.flatten()

for i, (coords, color) in enumerate(zip(query_points, colors)):
    # Quick visualization for each point
    results = visualize_similarity_results(
        input_image=test_image_path,
        query_coords=coords,
        overlay=True,
        alpha=0.6,
        show_query_point=True,
    )

    # Close the generated figure and create our own subplot
    plt.close(results["visualization"])

    # Get the similarity data and create our own visualization
    similarity_data = results["image_dict"]["image"][0]
    data, _ = processor.load_image(test_image_path)
    display_img = np.transpose(data, (1, 2, 0))

    # Show on subplot
    axes[i].imshow(display_img / 255.0)
    im = axes[i].imshow(similarity_data, cmap="turbo", alpha=0.6, vmin=0, vmax=1)
    axes[i].plot(
        coords[0],
        coords[1],
        "r*",
        markersize=15,
        markeredgecolor="white",
        markeredgewidth=2,
    )
    axes[i].set_title(f"Query: {color} Circle at {coords}")
    axes[i].axis("off")

    # Add colorbar
    plt.colorbar(im, ax=axes[i], fraction=0.046, pad=0.04)

plt.tight_layout()
plt.show()

# Test multiple query points query_points = [ (128, 128), # Red circle (384, 128), # Green circle (128, 384), # Blue circle (384, 384), # Yellow circle ] colors = ["Red", "Green", "Blue", "Yellow"] fig, axes = plt.subplots(2, 2, figsize=(20, 16)) axes = axes.flatten() for i, (coords, color) in enumerate(zip(query_points, colors)): # Quick visualization for each point results = visualize_similarity_results( input_image=test_image_path, query_coords=coords, overlay=True, alpha=0.6, show_query_point=True, ) # Close the generated figure and create our own subplot plt.close(results["visualization"]) # Get the similarity data and create our own visualization similarity_data = results["image_dict"]["image"][0] data, _ = processor.load_image(test_image_path) display_img = np.transpose(data, (1, 2, 0)) # Show on subplot axes[i].imshow(display_img / 255.0) im = axes[i].imshow(similarity_data, cmap="turbo", alpha=0.6, vmin=0, vmax=1) axes[i].plot( coords[0], coords[1], "r*", markersize=15, markeredgecolor="white", markeredgewidth=2, ) axes[i].set_title(f"Query: {color} Circle at {coords}") axes[i].axis("off") # Add colorbar plt.colorbar(im, ax=axes[i], fraction=0.046, pad=0.04) plt.tight_layout() plt.show()