app.py

import gradio as gr
import kornia as K
import kornia.feature as KF
import torch
import matplotlib

matplotlib.use("Agg")
import numpy as np
from plot_utils import plot_images, plot_lines, plot_color_line_matches

sold2 = KF.SOLD2(pretrained=True, config=None)
ransac = K.geometry.RANSAC(model_type="homography_from_linesegments", inl_th=3.0)


def infer(img1, img2, line_style: str):
    torch_img1 = K.image_to_tensor(img1).float() / 255.0
    torch_img2 = K.image_to_tensor(img2).float() / 255.0

    torch_img1_gray = K.color.rgb_to_grayscale(torch_img1)
    torch_img2_gray = K.color.rgb_to_grayscale(torch_img2)

    imgs = torch.stack(
        [torch_img1_gray, torch_img2_gray],
    )

    with torch.inference_mode():
        outputs = sold2(imgs)

    line_seg1 = outputs["line_segments"][0]
    line_seg2 = outputs["line_segments"][1]
    desc1 = outputs["dense_desc"][0]
    desc2 = outputs["dense_desc"][1]

    with torch.inference_mode():
        matches = sold2.match(line_seg1, line_seg2, desc1[None], desc2[None])

    valid_matches = matches != -1
    match_indices = matches[valid_matches]

    matched_lines1 = line_seg1[valid_matches]
    matched_lines2 = line_seg2[match_indices]

    imgs_to_plot = [K.tensor_to_image(torch_img1), K.tensor_to_image(torch_img2)]

    fig = plot_images(
        imgs_to_plot, ["Image 1 - detected lines", "Image 2 - detected lines"]
    )
    if line_style == "Line Matches":
        lines_to_plot = [line_seg1.numpy(), line_seg2.numpy()]
        plot_lines(lines_to_plot, fig, ps=3, lw=2, indices={0, 1})
    elif line_style == "Color Line Matches":
        plot_color_line_matches([matched_lines1, matched_lines2], fig, lw=2)
    elif line_style == "Line Segment Homography Warping":
        _, _, img1_warp_to2 = get_homography_values(
            matched_lines1, matched_lines2, torch_img1
        )
        fig = plot_images(
            [K.tensor_to_image(torch_img2), K.tensor_to_image(img1_warp_to2)],
            ["Image 2", "Image 1 wrapped to 2"],
        )
    elif line_style == "Matched Lines for Homography Warping":
        _, correspondence_mask, _ = get_homography_values(
            matched_lines1, matched_lines2, torch_img1
        )
        plot_color_line_matches(
            [matched_lines1[correspondence_mask], matched_lines2[correspondence_mask]],
            fig,
            lw=2,
        )
    return fig


def get_homography_values(matched_lines1, matched_lines2, torch_img1):
    H_ransac, correspondence_mask = ransac(
        matched_lines1.flip(dims=(2,)), matched_lines2.flip(dims=(2,))
    )
    img1_warp_to2 = K.geometry.warp_perspective(
        torch_img1[None], H_ransac[None], (torch_img1.shape[1:])
    )

    return H_ransac, correspondence_mask, img1_warp_to2


description = """In this space you can try out Line Detection and Segment Matching with the Kornia library as seen in [this tutorial](https://kornia-tutorials.readthedocs.io/en/latest/line_detection_and_matching_sold2.html).

Just upload two images of a scene with different view points, choose an option for output and run the demo.
"""


Iface = gr.Interface(
    fn=infer,
    inputs=[
        gr.components.Image(),
        gr.components.Image(),
        gr.components.Dropdown(
            [
                "Line Matches",
                "Color Line Matches",
                "Line Segment Homography Warping",
                "Matched Lines for Homography Warping",
            ],
            value="Line Matches",
            label="Options",
        ),
    ],
    outputs=gr.components.Plot(),
    examples=[["terrace0.JPG", "terrace1.JPG", "Line Matches"]],
    title="Line Segment Matching with Kornia",
    description=description,
).launch()