app.py

import os
os.system(f"pip install git+https://github.com/jinlinyi/PerspectiveFields.git@dev#egg=perspective2d")


import gradio as gr
import cv2
import copy
import torch
from PIL import Image, ImageDraw
from glob import glob
import numpy as np
import os.path as osp
from detectron2.config import get_cfg
from detectron2.data.detection_utils import read_image
from perspective2d.utils.predictor import VisualizationDemo
import perspective2d.modeling  # noqa
from perspective2d.config import get_perspective2d_cfg_defaults
from perspective2d.utils import draw_from_r_p_f_cx_cy




title = "Perspective Fields Demo"

description = """
<p style="text-align: center">
    <a href="https://jinlinyi.github.io/PerspectiveFields/" target="_blank">Project Page</a> | 
    <a href="https://arxiv.org/abs/2212.03239" target="_blank">Paper</a> | 
    <a href="https://github.com/jinlinyi/PerspectiveFields" target="_blank">Code</a> | 
    <a href="https://www.youtube.com/watch?v=sN5B_ZvMva8&themeRefresh=1" target="_blank">Video</a>
</p>
<h2>Gradio Demo</h2>
<p>Try our Gradio demo for Perspective Fields for single image camera calibration. You can click on one of the provided examples or upload your own image.</p>
<h3>Available Models:</h3>
<ol>
    <li><strong>PersNet-360Cities:</strong> PerspectiveNet trained on the 360Cities dataset. This model predicts perspective fields and is designed to be robust and generalize well to both indoor and outdoor images.</li>
    <li><strong>PersNet_Paramnet-GSV-uncentered:</strong> A combination of PerspectiveNet and ParamNet trained on the Google Street View (GSV) dataset. This model predicts camera Roll, Pitch, and Field of View (FoV), as well as the Principal Point location.</li>
    <li><strong>PersNet_Paramnet-GSV-centered:</strong> PerspectiveNet+ParamNet trained on the GSV dataset. This model assumes the principal point is at the center of the image and predicts camera Roll, Pitch, and FoV.</li>
</ol>
"""


article = """
<p style='text-align: center'><a href='https://arxiv.org/abs/2212.03239' target='_blank'>Perspective Fields for Single Image Camera Calibrations</a> | <a href='https://github.com/jinlinyi/PerspectiveFields' target='_blank'>Github Repo</a></p>
"""




def setup_cfg(args):
    cfgs = []
    configs = args['config_file'].split('#')
    weights_id = args['opts'].index('MODEL.WEIGHTS') + 1
    weights = args['opts'][weights_id].split('#')
    for i, conf in enumerate(configs):
        if len(conf) != 0:
            tmp_opts = copy.deepcopy(args['opts'])
            tmp_opts[weights_id] = weights[i]
            cfg = get_cfg()
            get_perspective2d_cfg_defaults(cfg)
            cfg.merge_from_file(conf)
            cfg.merge_from_list(tmp_opts)
            cfg.freeze()
            cfgs.append(cfg)
    return cfgs

def resize_fix_aspect_ratio(img, field, target_width=None, target_height=None):
    height = img.shape[0]
    width = img.shape[1]
    if target_height is None:
        factor = target_width / width
    elif target_width is None:
        factor = target_height / height
    else:
        factor = max(target_width / width, target_height / height)
    if factor == target_width / width:
        target_height = int(height * factor)
    else:
        target_width = int(width * factor)
        
    img = cv2.resize(img, (target_width, target_height))
    for key in field:
        if key not in ['up', 'lati']:
            continue
        tmp = field[key].numpy()
        transpose = len(tmp.shape) == 3
        if transpose:
            tmp = tmp.transpose(1,2,0)
        tmp = cv2.resize(tmp, (target_width, target_height))
        if transpose:
            tmp = tmp.transpose(2,0,1)
        field[key] = torch.tensor(tmp)
    return img, field


def inference(img, model_type):
    if model_type is None:
        return None, ""
    perspective_cfg_list = setup_cfg(model_zoo[model_type])
    demo = VisualizationDemo(cfg_list=perspective_cfg_list)

    # img = read_image(image_path, format="BGR")
    img = img[..., ::-1] # rgb->bgr
    pred = demo.run_on_image(img)
    field = {
        'up': pred['pred_gravity_original'].cpu().detach(),
        'lati': pred['pred_latitude_original'].cpu().detach(),
    }
    img, field = resize_fix_aspect_ratio(img, field, 640)
    if not model_zoo[model_type]['param']:
        pred_vis = demo.draw(
            image=img,
            latimap=field['lati'],
            gravity=field['up'],
            latimap_format=pred['pred_latitude_original_mode'],
        ).get_image()
        param = "Not Implemented"
    else:
        if 'pred_general_vfov' not in pred.keys():
            pred['pred_general_vfov'] = pred['pred_vfov']
        if 'pred_rel_cx' not in pred.keys():
            pred['pred_rel_cx'] = torch.FloatTensor([0])
        if 'pred_rel_cy' not in pred.keys():
            pred['pred_rel_cy'] = torch.FloatTensor([0])
            
        r_p_f_rad = np.radians(
            [
                pred['pred_roll'].cpu().item(),
                pred['pred_pitch'].cpu().item(),
                pred['pred_general_vfov'].cpu().item(),
            ]
        )
        cx_cy = [
            pred['pred_rel_cx'].cpu().item(),
            pred['pred_rel_cy'].cpu().item(),
        ]
        param = f"roll {pred['pred_roll'].cpu().item() :.2f}\npitch {pred['pred_pitch'].cpu().item() :.2f}\nfov {pred['pred_general_vfov'].cpu().item() :.2f}\n"
        param += f"principal point {pred['pred_rel_cx'].cpu().item() :.2f} {pred['pred_rel_cy'].cpu().item() :.2f}"
        pred_vis = draw_from_r_p_f_cx_cy(
            img[:,:,::-1], 
            *r_p_f_rad,
            *cx_cy,
            'rad',
            up_color=(0,1,0),
        )
    return Image.fromarray(pred_vis), param

examples = []
for img_name in glob('assets/imgs/*.*g'):
    examples.append([img_name])
print(examples)

device = 'cuda' if torch.cuda.is_available() else 'cpu'
model_zoo = {
    'PersNet-360Cities': {
        'weights': ['https://www.dropbox.com/s/czqrepqe7x70b7y/cvpr2023.pth'],
        'opts': ['MODEL.WEIGHTS', 'models/cvpr2023.pth', 'MODEL.DEVICE', device,],
        'config_file': 'models/cvpr2023.yaml',
        'param': False,
    },
    'PersNet_Paramnet-GSV-uncentered': {
        'weights': ['https://www.dropbox.com/s/ufdadxigewakzlz/paramnet_gsv_rpfpp.pth'],
        'opts': ['MODEL.WEIGHTS', 'models/paramnet_gsv_rpfpp.pth', 'MODEL.DEVICE', device,],
        'config_file': 'models/paramnet_gsv_rpfpp.yaml',
        'param': True,
    },
    # trained on GSV dataset, predicts Perspective Fields + camera parameters (roll, pitch, fov), assuming centered principal point
    'PersNet_Paramnet-GSV-centered': {
        'weights': ['https://www.dropbox.com/s/g6xwbgnkggapyeu/paramnet_gsv_rpf.pth'],
        'opts': ['MODEL.WEIGHTS', 'models/paramnet_gsv_rpf.pth', 'MODEL.DEVICE', device,],
        'config_file': 'models/paramnet_gsv_rpf.yaml',
        'param': True,
    },
}
for model_id in model_zoo:
    html = model_zoo[model_id]['weights'][0]
    if not os.path.exists(os.path.join('models', html.split('/')[-1])):
        os.system(f"wget -P models/ {html}")

info = """Select model\n"""
gr.Interface(
    fn=inference,
    inputs=[
        "image", 
        gr.Radio(
            list(model_zoo.keys()), 
            value=list(sorted(model_zoo.keys()))[0], 
            label="Model", 
            info=info,
        ),
    ],
    outputs=[gr.Image(label='Perspective Fields'), gr.Textbox(label='Pred Camera Parameters')],
    title=title,
    description=description,
    article=article,
    examples=examples,
).launch()