utils/depth_estimation.py

# utils/depth_estimation.py

import torch
import numpy as np
from PIL import Image
import open3d as o3d
from transformers import DPTImageProcessor, DPTForDepthEstimation
from pathlib import Path
import logging
logging.getLogger("transformers.modeling_utils").setLevel(logging.ERROR)
from utils.image_utils import (
    resize_image_with_aspect_ratio
)

# Load models once during module import
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-large")
depth_model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large", ignore_mismatched_sizes=True)

def estimate_depth(image):
    # Ensure image is in RGB mode
    if image.mode != "RGB":
        image = image.convert("RGB")

    # Resize the image for the model
    image_resized = image.resize(
        (image.width, image.height),
        Image.Resampling.LANCZOS
    )

    # Prepare image for the model
    encoding = image_processor(image_resized, return_tensors="pt")

    # Forward pass
    with torch.no_grad():
        outputs = depth_model(**encoding)
        predicted_depth = outputs.predicted_depth

    # Interpolate to original size
    prediction = torch.nn.functional.interpolate(
        predicted_depth.unsqueeze(1),
        size=(image.height, image.width),
        mode="bicubic",
        align_corners=False,
    ).squeeze()

    # Convert to depth image
    output = prediction.cpu().numpy()
    depth_min = output.min()
    depth_max = output.max()
    max_val = (2**8) - 1

    # Normalize and convert to 8-bit image
    depth_image = max_val * (output - depth_min) / (depth_max - depth_min)
    depth_image = depth_image.astype("uint8")

    depth_pil = Image.fromarray(depth_image)

    return depth_pil, output

def create_3d_model(rgb_image, depth_array, voxel_size_factor=0.01):
    depth_o3d = o3d.geometry.Image(depth_array.astype(np.float32))
    rgb_o3d = o3d.geometry.Image(np.array(rgb_image))

    rgbd_image = o3d.geometry.RGBDImage.create_from_color_and_depth(
        rgb_o3d,
        depth_o3d,
        convert_rgb_to_intensity=False
    )

    # Create a point cloud from the RGBD image
    camera_intrinsic = o3d.camera.PinholeCameraIntrinsic(
        rgb_image.width,
        rgb_image.height,
        fx=1.0,
        fy=1.0,
        cx=rgb_image.width / 2.0,
        cy=rgb_image.height / 2.0,
    )

    pcd = o3d.geometry.PointCloud.create_from_rgbd_image(
        rgbd_image,
        camera_intrinsic
    )

    # Voxel downsample
    voxel_size = max(pcd.get_max_bound() - pcd.get_min_bound()) * voxel_size_factor
    voxel_grid = o3d.geometry.VoxelGrid.create_from_point_cloud(pcd, voxel_size=voxel_size)

    # Save the 3D model to a temporary file
    temp_dir = Path.cwd() / "temp_models"
    temp_dir.mkdir(exist_ok=True)
    model_path = temp_dir / "model.ply"
    o3d.io.write_voxel_grid(str(model_path), voxel_grid)

    return str(model_path)

def generate_depth_and_3d(input_image_path, voxel_size_factor):
    image = Image.open(input_image_path).convert("RGB")
    resized_image = resize_image_with_aspect_ratio(image, 2688, 1680)
    depth_image, depth_array = estimate_depth(resized_image)
    model_path = create_3d_model(resized_image, depth_array, voxel_size_factor=voxel_size_factor)
    return depth_image, model_path

def generate_depth_button_click(depth_image_source, voxel_size_factor, input_image, output_image, overlay_image, bordered_image_output):
    if depth_image_source == "Input Image":
        image_path = input_image
    elif depth_image_source == "Output Image":
        image_path = output_image
    elif depth_image_source == "Image with Margins":
        image_path = bordered_image_output
    else:
        image_path = overlay_image

    return generate_depth_and_3d(image_path, voxel_size_factor)