app.py

import gradio as gr
import numpy as np
import torch
import os
import cv2
from PIL import Image
import torchvision.transforms as transforms
from net.dornet import Net
from net.dornet_ddp import Net_ddp

# init
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
net = Net(tiny_model=False).to(device)
model_ckpt_map = {
    "RGB-D-D": "./checkpoints/RGBDD.pth",
    "TOFDSR": "./checkpoints/TOFDSR.pth"
}

# load model
def load_model(model_type: str):
    global net
    ckpt_path = model_ckpt_map[model_type]
    print(f"Loading weights from: {ckpt_path}")
    if model_type == "RGB-D-D":
        net = Net(tiny_model=False).to(device)
    elif model_type == "TOFDSR":
        net = Net_ddp(tiny_model=False).srn.to(device)
    else:
        raise ValueError(f"Unknown model_type: {model_type}")

    net.load_state_dict(torch.load(ckpt_path, map_location=device))
    net.eval()

load_model("RGB-D-D")


# data process
def preprocess_inputs(rgb_image: Image.Image, lr_depth: Image.Image):
    image = np.array(rgb_image.convert("RGB")).astype(np.float32)
    h, w, _ = image.shape
    lr = np.array(lr_depth.resize((w, h), Image.BICUBIC)).astype(np.float32)
    # Normalize depth
    max_out, min_out = 5000.0, 0.0
    lr = (lr - min_out) / (max_out - min_out)
    # Normalize RGB
    maxx, minn = np.max(image), np.min(image)
    image = (image - minn) / (maxx - minn)
    # To tensor
    data_transform = transforms.Compose([transforms.ToTensor()])
    image = data_transform(image).float()
    lr = data_transform(np.expand_dims(lr, 2)).float()
    # Add batch dimension
    lr = lr.unsqueeze(0).to(device)
    image = image.unsqueeze(0).to(device)
    return image, lr, min_out, max_out


# model inference
@torch.no_grad()
def infer(rgb_image: Image.Image, lr_depth: Image.Image, model_type: str):
    load_model(model_type)  # reset weight

    image, lr, min_out, max_out = preprocess_inputs(rgb_image, lr_depth)

    if model_type == "RGB-D-D":
        out = net(x_query=lr, rgb=image)
    elif model_type == "TOFDSR":
        out, _ = net(x_query=lr, rgb=image)

    pred = out[0, 0] * (max_out - min_out) + min_out
    pred = pred.cpu().numpy().astype(np.uint16)
    # raw
    pred_gray = Image.fromarray(pred)

    # heat
    pred_norm = (pred - np.min(pred)) / (np.max(pred) - np.min(pred)) * 255
    pred_vis = pred_norm.astype(np.uint8)
    pred_heat = cv2.applyColorMap(pred_vis, cv2.COLORMAP_PLASMA)
    pred_heat = cv2.cvtColor(pred_heat, cv2.COLOR_BGR2RGB)
    return pred_gray, Image.fromarray(pred_heat)


# Gradio
# demo = gr.Interface(
#     fn=infer,
#     inputs=[
#         gr.Image(label="RGB Image", type="pil"),
#         gr.Image(label="Low-res Depth", type="pil", image_mode="I"),
#         gr.Dropdown(choices=["RGB-D-D", "TOFDSR"], label="Model Type", value="RGB-D-D")
#     ],
#     outputs=[
#         gr.Image(label="DORNet Output", type="pil", elem_classes=["output-image"]),
#         gr.Image(label="Normalized Output", type="pil", elem_classes=["output-image"])
#     ],
#     examples=[
#         ["examples/RGB-D-D/20200518160957_RGB.jpg", "examples/RGB-D-D/20200518160957_LR_fill_depth.png", "RGB-D-D"],
#         ["examples/TOFDSR/2020_09_08_13_59_59_435_rgb_rgb_crop.png", "examples/TOFDSR/2020_09_08_13_59_59_435_rgb_depth_crop_fill.png", "TOFDSR"],
#     ],
#     allow_flagging="never",
#     title="DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution \n CVPR 2025 (Oral Presentation)",
#     css="""
#         .output-image {
#             display: flex;
#             justify-content: center;
#             align-items: center;
#         }
#         .output-image img {
#             margin: auto;
#             display: block;
#         }
#         """
# )
#
# demo.launch(share=True)
Intro = """
## DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution
[📄 Paper](https://arxiv.org/pdf/2410.11666) • [💻 Code](https://github.com/yanzq95/DORNet) • [📦 Model](https://huggingface.co/wzxwyx/DORNet/tree/main)
"""

with gr.Blocks(css="""
    .output-image {
        display: flex;
        justify-content: center;
        align-items: center;
    }
    .output-image img {
        margin: auto;
        display: block;
    }
""") as demo:
    gr.Markdown(Intro)

    with gr.Row():
        with gr.Column():
            rgb_input = gr.Image(label="RGB Image", type="pil")
            lr_input = gr.Image(label="Low-res Depth", type="pil", image_mode="I")
        with gr.Column():
            output1 = gr.Image(label="DORNet Output", type="pil", elem_classes=["output-image"])
            output2 = gr.Image(label="Normalized Output", type="pil", elem_classes=["output-image"])

    model_selector = gr.Dropdown(choices=["RGB-D-D", "TOFDSR"], label="Model Type", value="RGB-D-D")
    run_button = gr.Button("Run Inference")

    gr.Examples(
        examples=[
            ["examples/RGB-D-D/20200518160957_RGB.jpg", "examples/RGB-D-D/20200518160957_LR_fill_depth.png", "RGB-D-D"],
            ["examples/TOFDSR/2020_09_08_13_59_59_435_rgb_rgb_crop.png", "examples/TOFDSR/2020_09_08_13_59_59_435_rgb_depth_crop_fill.png", "TOFDSR"],
        ],
        inputs=[rgb_input, lr_input, model_selector],
        outputs=[output1, output2],
        label="Try Examples ↓"
    )

    run_button.click(fn=infer, inputs=[rgb_input, lr_input, model_selector], outputs=[output1, output2])

demo.launch(share=True)