app.py

import time

import gradio as gr
import matplotlib.cm as cm
import numpy as np
import plotly.graph_objects as go
import spaces
import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForKeypointMatching
from transformers.image_utils import to_numpy_array


@spaces.GPU
def process_images(image1, image2, model_name):
    """
    Process two images and return a plot of the matching keypoints.
    """
    if image1 is None or image2 is None:
        return None

    images = [image1, image2]

    processor = AutoImageProcessor.from_pretrained(model_name)
    model = AutoModelForKeypointMatching.from_pretrained(model_name, device_map="auto")
    inputs = processor(images, return_tensors="pt")
    inputs = inputs.to(model.device)
    print(
        f"Model {model_name} is on device: {model.device} and inputs are on device: {inputs['pixel_values'].device}"
    )

    with torch.no_grad():
        outputs = model(**inputs)

    image_sizes = [[(image.height, image.width) for image in images]]
    outputs = processor.post_process_keypoint_matching(
        outputs, image_sizes, threshold=0.2
    )
    output = outputs[0]

    image1 = to_numpy_array(image1)
    image2 = to_numpy_array(image2)

    height0, width0 = image1.shape[:2]
    height1, width1 = image2.shape[:2]

    # Create PIL image from numpy array
    pil_img = Image.fromarray((image1 / 255.0 * 255).astype(np.uint8))
    pil_img2 = Image.fromarray((image2 / 255.0 * 255).astype(np.uint8))

    fig = go.Figure()

    # Create colormap (red-yellow-green: red for low scores, green for high scores)
    colormap = cm.RdYlGn

    # Get keypoints
    keypoints0_x, keypoints0_y = output["keypoints0"].unbind(1)
    keypoints1_x, keypoints1_y = output["keypoints1"].unbind(1)

    # Add a separate trace for each match (line + markers) to enable highlighting
    for keypoint0_x, keypoint0_y, keypoint1_x, keypoint1_y, matching_score in zip(
        keypoints0_x,
        keypoints0_y,
        keypoints1_x,
        keypoints1_y,
        output["matching_scores"],
    ):
        color_val = matching_score.item()
        rgba_color = colormap(color_val)

        # Convert to rgba string with transparency
        color = f"rgba({int(rgba_color[0] * 255)}, {int(rgba_color[1] * 255)}, {int(rgba_color[2] * 255)}, 0.8)"

        hover_text = (
            f"Score: {matching_score.item():.3f}<br>"
            f"Point 1: ({keypoint0_x.item():.1f}, {keypoint0_y.item():.1f})<br>"
            f"Point 2: ({keypoint1_x.item():.1f}, {keypoint1_y.item():.1f})"
        )

        fig.add_trace(
            go.Scatter(
                x=[keypoint0_x.item(), keypoint1_x.item() + width0],
                y=[keypoint0_y.item(), keypoint1_y.item()],
                mode="lines+markers",
                line=dict(color=color, width=2),
                marker=dict(color=color, size=5, opacity=0.8),
                hoverinfo="text",
                hovertext=hover_text,
                showlegend=False,
            )
        )

    # Update layout to use images as background
    fig.update_layout(
        xaxis=dict(
            range=[0, width0 + width1],
            showgrid=False,
            zeroline=False,
            showticklabels=False,
        ),
        yaxis=dict(
            range=[max(height0, height1), 0],
            showgrid=False,
            zeroline=False,
            showticklabels=False,
            scaleanchor="x",
            scaleratio=1,
        ),
        margin=dict(l=0, r=0, t=0, b=0),
        autosize=True,
        images=[
            dict(
                source=pil_img,
                xref="x",
                yref="y",
                x=0,
                y=0,
                sizex=width0,
                sizey=height0,
                sizing="stretch",
                opacity=1,
                layer="below",
            ),
            dict(
                source=pil_img2,
                xref="x",
                yref="y",
                x=width0,
                y=0,
                sizex=width1,
                sizey=height1,
                sizing="stretch",
                opacity=1,
                layer="below",
            ),
        ],
    )

    return fig


# Create the Gradio interface
with gr.Blocks(title="EfficientLoFTR Matching Demo") as demo:
    gr.Markdown("# EfficientLoFTR Matching Demo")
    gr.Markdown(
        "Upload two images and get a side-by-side matching of your images using EfficientLoFTR."
    )
    gr.Markdown("""
        ## How to use:
        1. Select an EfficientLoFTR model (Original EfficientLoFTR or MatchAnything)
        2. Upload two images using the file uploaders below
        3. Click the 'Match Images' button
        4. View the matched output image below. Higher scores are green, lower scores are red.
        
        The app will create a side-by-side matching of your images using EfficientLoFTR.
        You can also select an example image pair from the dataset below.
    """)

    with gr.Row():
        # Detector choice selector
        detector_choice = gr.Radio(
            choices=[("Original EfficientLoFTR", "zju-community/efficientloftr"), ("MatchAnything", "zju-community/matchanything_eloftr")],
            value="Original EfficientLoFTR",
            label="EfficientLoFTR Model",
            info="Choose between original EfficientLoFTR or MatchAnything"
        )

    with gr.Row():
        # Input images on the same row
        image1 = gr.Image(label="First Image", type="pil")
        image2 = gr.Image(label="Second Image", type="pil")

    # Process button
    process_btn = gr.Button("Match Images", variant="primary")

    # Output plot
    output_plot = gr.Plot(label="Matching Results", scale=2)

    # Connect the function
    process_btn.click(fn=process_images, inputs=[image1, image2, detector_choice], outputs=[output_plot])

    # Add some example usage

    examples = gr.Dataset(
        components=[image1, image2],
        label="Example Image Pairs",
        samples_per_page=100,
        samples=[
            [
                "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/united_states_capitol_98169888_3347710852.jpg",
                "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/united_states_capitol_26757027_6717084061.jpg",
            ],
            [
                "https://raw.githubusercontent.com/cvg/LightGlue/refs/heads/main/assets/DSC_0410.JPG",
                "https://raw.githubusercontent.com/cvg/LightGlue/refs/heads/main/assets/DSC_0411.JPG",
            ],
            [
                "https://raw.githubusercontent.com/cvg/LightGlue/refs/heads/main/assets/sacre_coeur1.jpg",
                "https://raw.githubusercontent.com/cvg/LightGlue/refs/heads/main/assets/sacre_coeur2.jpg",
            ],
            [
                "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/piazza_san_marco_06795901_3725050516.jpg",
                "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/piazza_san_marco_58751010_4849458397.jpg",
            ],
            [
                "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/london_bridge_19481797_2295892421.jpg",
                "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/london_bridge_78916675_4568141288.jpg",
            ],
            # MatchAnything multi-modality pairs
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/MTV_thermal_vis_1.jpg",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/MTV_thermal_vis_2.jpg",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/MTV_thermal_vis_pair2_1.jpg",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/MTV_thermal_vis_pair2_2.jpg",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/ct_mr_1.png",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/ct_mr_2.png",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/mri_ut_1.jpg",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/mri_ut_2.jpg",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/robot_render_1.png",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/robot_real_world_2.png",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/thermal_vis_1.jpg",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/thermal_vis_2.jpg",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_event_1.png",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_event_2.png",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_map_1.jpg",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_map_2.jpg",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_map_pair2_1.jpg",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_map_pair2_2.jpg",
            ],
            [
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_thermal_ground_1.png",
                "https://huggingface.co/spaces/LittleFrog/MatchAnything/resolve/main/imcui/datasets/multi_modality_pairs/vis_thermal_ground_2.png",
            ],
        ],
    )

    examples.select(lambda x: (x[0], x[1]), [examples], [image1, image2])

if __name__ == "__main__":
    demo.launch()