app.py

from email.policy import default
import gradio as gr
from transformers import DPTFeatureExtractor, DPTForDepthEstimation
import torch
import numpy as np
from PIL import Image
import open3d as o3d
from pathlib import Path
import os

feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large")


def process_image(image_path, voxel_s):
    voxel_s = max(voxel_s/500, 0.0001)
    image_path = Path(image_path)
    image_raw = Image.open(image_path)
    image = image_raw.resize(
        (800, int(800 * image_raw.size[1] / image_raw.size[0])),
        Image.Resampling.LANCZOS)

    # prepare image for the model
    encoding = feature_extractor(image, return_tensors="pt")

    # forward pass
    with torch.no_grad():
        outputs = model(**encoding)
        predicted_depth = outputs.predicted_depth

    # interpolate to original size
    prediction = torch.nn.functional.interpolate(
        predicted_depth.unsqueeze(1),
        size=image.size[::-1],
        mode="bicubic",
        align_corners=False,
    ).squeeze()
    output = prediction.cpu().numpy()
    depth_image = (output * 255 / np.max(output)).astype('uint8')
    try:
        gltf_path = create_3d_voxels_obj(
            np.array(image), depth_image, image_path, voxel_s)
        img = Image.fromarray(depth_image)
        return [img, gltf_path, gltf_path]
    except Exception as e:
        print("Error reconstructing 3D model")
        raise Exception("Error reconstructing 3D model")


def create_3d_voxels_obj(rgb_image, depth_image, image_path, voxel_s):
    depth_o3d = o3d.geometry.Image(depth_image)
    image_o3d = o3d.geometry.Image(rgb_image)
    rgbd_image = o3d.geometry.RGBDImage.create_from_color_and_depth(
        image_o3d, depth_o3d, convert_rgb_to_intensity=False)
    w = int(depth_image.shape[1])
    h = int(depth_image.shape[0])

    camera_intrinsic = o3d.camera.PinholeCameraIntrinsic()
    camera_intrinsic.set_intrinsics(w, h, 500, 500, w/2, h/2)

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

    print('normals')
    pcd.normals = o3d.utility.Vector3dVector(
        np.zeros((1, 3)))  # invalidate existing normals
    pcd.estimate_normals(
        search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.01, max_nn=30))
    pcd.orient_normals_towards_camera_location(
        camera_location=np.array([0., 0., 1000.]))
    pcd.transform([[1, 0, 0, 0],
                   [0, -1, 0, 0],
                   [0, 0, -1, 0],
                   [0, 0, 0, 1]])
    pcd.transform([[-1, 0, 0, 0],
                   [0, 1, 0, 0],
                   [0, 0, 1, 0],
                   [0, 0, 0, 1]])

    print('voxels')

    # ref https://towardsdatascience.com/how-to-automate-voxel-modelling-of-3d-point-cloud-with-python-459f4d43a227
    voxel_size = round(
        max(pcd.get_max_bound()-pcd.get_min_bound())*voxel_s, 10)
    print("Voxel size", voxel_size, "voxel_s", voxel_s)
    voxel_grid = o3d.geometry.VoxelGrid.create_from_point_cloud(
        pcd, voxel_size=voxel_size)
    voxels = voxel_grid.get_voxels()

    vox_mesh = o3d.geometry.TriangleMesh()
    for v in voxels:
        cube = o3d.geometry.TriangleMesh.create_box(width=1, height=1, depth=1)
        cube.paint_uniform_color(v.color)
        cube.translate(v.grid_index, relative=False)
        vox_mesh += cube
    print(voxel_grid, vox_mesh)

    gltf_path = f'./{image_path.stem}.gltf'
    o3d.io.write_triangle_mesh(gltf_path, vox_mesh, write_triangle_uvs=True)
    return gltf_path


title = "Demo: zero-shot depth estimation with DPT + 3D Voxels reconstruction"
description = "This demo is a variation from the original <a href='https://huggingface.co/spaces/nielsr/dpt-depth-estimation' target='_blank'>DPT Demo</a>. It uses the DPT model to predict the depth of an image and then reconstruct the 3D model as voxels."
examples = [["examples/" + img, 10] for img in os.listdir("examples/")]

iface = gr.Interface(fn=process_image,
                     inputs=[
                         gr.Image(
                             type="filepath", label="Input Image"),
                         gr.Slider(value=10, minimum=5, maximum=100, step=1, label="Voxel Size",)
                     ],
                     outputs=[
                         gr.Image(label="predicted depth", type="pil"),
                         gr.Model3D(label="3d mesh reconstruction", clear_color=[
                             1.0, 1.0, 1.0, 1.0]),
                         gr.File(label="3d gLTF")
                     ],
                     title=title,
                     description=description,
                     examples=examples,
                     allow_flagging="never",
                     cache_examples=False)
iface.launch(debug=True)