app.py

# Copyright 2022-present NAVER Corp.
# CC BY-NC-SA 4.0
# Available only for non-commercial use

import gradio as gr
import sys, os
import torch
import matplotlib.pylab as plt
from PIL import ImageOps

def pump_matching(img1, img2, trained_with_st=False, scale=300, max_scale=1, max_rot=0, use_gpu=False):

    img1 = ImageOps.exif_transpose(img1)
    img2 = ImageOps.exif_transpose(img2)

    use_singlescale = max_scale==1 and max_rot==0
    if use_singlescale: # single 
        from test_singlescale import Main, arg_parser
    else:
        from test_multiscale import Main, arg_parser
    parser = arg_parser()

    args_list = ['--img1','dummy','--img2','dummy','--post-filter', '--desc','PUMP-stytrf' if trained_with_st else 'PUMP','--resize',str(scale)]
    if not use_gpu:
        args_list += ['--device', 'cpu']
    if not use_singlescale:
        args_list += ['--max-scale',str(max_scale),'--max-rot',str(max_rot)]
        
    args = parser.parse_args(args_list)
    
    corres = Main().run_from_args_with_images(img1, img2, args)
        
    fig1 = plt.figure(1)
    plt.clf()
    ax1 = plt.gca()
    ax1.imshow(img1)
    ax1.axis('off')
    plt.tight_layout(pad=0)  
    
    fig2 = plt.figure(2)
    plt.clf()
    ax2 = plt.gca()
    ax2.imshow(img2)
    ax2.axis('off')
    plt.tight_layout(pad=0) 
       
    from tools.viz import plot_grid
    if corres.shape[-1] > 4:
        corres = corres[corres[:,4]>0,:] # select non-null correspondences
    if corres.shape[0]>0: plot_grid(corres, ax1, ax2, marker='+')

    img1 = None 
    img2 = None

    return fig1, fig2

has_cuda = torch.cuda.is_available() and torch.cuda.device_count()>0

title = "PUMP local descriptor demo"
description = "This is a visualization demo for the PUMP local descriptors presented in our CVPR 2022 paper <b><a href='https://europe.naverlabs.com/research/publications/pump-pyramidal-and-uniqueness-matching-priors-for-unsupervised-learning-of-local-features/' target='_blank'>PUMP: Pyramidal and Uniqueness Matching Priors for Unsupervised Learning of Local Features</a></b>.</p><p><b>WARNING:</b> this demo runs on cpus with downscaled images, without multi-scale or multi-rotations testing, due to limited memory and computational resources, please check out our <a href='https://github.com/naver/pump' target='_blank'>original github repo</a> for these features.</p>" 

article = "<p style='text-align: center'><a href='https://github.com/naver/pump' target='_blank'>Original Github Repo</a></p>"

iface = gr.Interface(
    fn=pump_matching,
    inputs=[
        gr.inputs.Image(shape=None, type="pil", label="First Image"),
        gr.inputs.Image(shape=None, type="pil", label="Second Image"),
        gr.inputs.Checkbox(default=False, label="Use the model trained with style transfer"),
        #gr.inputs.Slider(minimum=300, maximum=600, default=400, step=1, label="Original test scale"),
        #gr.inputs.Slider(minimum=1, maximum=4, default=1, step=0.1, label="Multi Scale Testing - maximum scale (makes it slower)"),
        #gr.inputs.Slider(minimum=0, maximum=180, default=0, step=45, label="Multi Rotation Testing - max rot (makes it slower)"),]
        #+ ([gr.inputs.Checkbox(default=True, label='Use GPU instead of CPU')] if has_cuda else []),"""
        ],
    outputs=[
        gr.outputs.Image(type="plot", label="Matches in the first image"),
        gr.outputs.Image(type="plot", label="Matches in the second image"),
        ],
    title=title,
    theme='peach',
    description=description,
    article=article,
    examples=[
        ['datasets/gradio_demo/cat_src.jpg','datasets/gradio_demo/cat_tgt.jpg',False],#,400,1,0]+([True] if has_cuda else []),
        ['datasets/gradio_demo/food_src.jpg','datasets/gradio_demo/food_tgt.jpg',False],#,400,1,0]+([True] if has_cuda else []),
        ['datasets/demo_warp/mountains_src.jpg','datasets/demo_warp/mountains_tgt.jpg',False],#,400,1,0]+([True] if has_cuda else []),
    ]
)
iface.launch(enable_queue=True)