Update to 2025 codebase, gradio

.gitignore

@@ -43,4 +43,5 @@ workspace.code-workspace
 # log files
 .pnpm-debug.log
 venv/
-*.db-journal
+*.db-journal
+/.vs

README.md

@@ -1,10 +1,10 @@
 ---
-title: Dpt Depth Estimation + 3D
+title: DPT Depth Estimation + 3D
 emoji: ⚡
 colorFrom: blue
 colorTo: red
 sdk: gradio
-sdk_version: 4.26.0
+sdk_version: 5.16.1
 app_file: app.py
 pinned: false
 short_description: Image to 3D with DPT + 3D Point Cloud

app.py

@@ -1,119 +1,202 @@
+import os
+from pathlib import Path
+
 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
+import torch
+from PIL import Image
+from transformers import DPTForDepthEstimation, DPTImageProcessor
 
-feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large")
+# Initialize the image processor and depth estimation model
+image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-large")
 model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large")
 
 
-def process_image(image_path):
+def process_image(image_path, resized_width=800, z_scale=208):
+    """
+    Processes the input image to generate a depth map and a 3D mesh reconstruction.
+
+    Args:
+        image_path (str): The file path to the input image.
+
+    Returns:
+        list: A list containing the depth image, 3D mesh reconstruction, and GLTF file path.
+    """
     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,
-    )
+    if not image_path.exists():
+        raise ValueError("Image file not found")
+
+    # Load and resize the image
+    image_raw = Image.open(image_path).convert("RGB")
+    print(f"Original size: {image_raw.size}")
+    resized_height = int(resized_width * image_raw.size[1] / image_raw.size[0])
+    image = image_raw.resize((resized_width, resized_height), Image.Resampling.LANCZOS)
+    print(f"Resized size: {image.size}")
 
-    # prepare image for the model
-    encoding = feature_extractor(image, return_tensors="pt")
+    # Prepare image for the model
+    encoding = image_processor(image, return_tensors="pt")
 
-    # forward pass
+    # Perform depth estimation
     with torch.no_grad():
         outputs = model(**encoding)
         predicted_depth = outputs.predicted_depth
 
-    # interpolate to original size
+    # Interpolate depth to match the image size
     prediction = torch.nn.functional.interpolate(
         predicted_depth.unsqueeze(1),
-        size=image.size[::-1],
+        size=(image.height, image.width),
         mode="bicubic",
-        align_corners=False,
+        align_corners=True,
     ).squeeze()
-    output = prediction.cpu().numpy()
-    depth_image = (output * 255 / np.max(output)).astype("uint8")
+
+    # Normalize the depth image to 8-bit
+    prediction = prediction.cpu().numpy()
+    depth_min, depth_max = prediction.min(), prediction.max()
+    depth_image = ((prediction - depth_min) / (depth_max - depth_min) * 255).astype("uint8")
+
     try:
-        gltf_path = create_3d_obj(np.array(image), depth_image, image_path)
-        img = Image.fromarray(depth_image)
-        return [img, gltf_path, gltf_path]
-    except Exception as e:
-        gltf_path = create_3d_obj(np.array(image), depth_image, image_path, depth=8)
-        img = Image.fromarray(depth_image)
-        return [img, gltf_path, gltf_path]
-    except:
-        print("Error reconstructing 3D model")
-        raise Exception("Error reconstructing 3D model")
-
-
-def create_3d_obj(rgb_image, depth_image, image_path, depth=10):
+        gltf_path = create_3d_obj(np.array(image), prediction, image_path, depth=10, z_scale=z_scale)
+    except Exception:
+        gltf_path = create_3d_obj(np.array(image), prediction, image_path, depth=8, z_scale=z_scale)
+
+    img = Image.fromarray(depth_image)
+    return [img, gltf_path, gltf_path]
+
+
+def create_3d_obj(rgb_image, raw_depth, image_path, depth=10, z_scale=200):
+    """
+    Creates a 3D object from RGB and depth images.
+
+    Args:
+        rgb_image (np.ndarray): The RGB image as a NumPy array.
+        raw_depth (np.ndarray): The raw depth data.
+        image_path (Path): The path to the original image.
+        depth (int, optional): Depth parameter for Poisson reconstruction. Defaults to 10.
+        z_scale (float, optional): Scaling factor for the Z-axis. Defaults to 200.
+
+    Returns:
+        str: The file path to the saved GLTF model.
+    """
+    # Normalize the depth image
+    depth_image = ((raw_depth - raw_depth.min()) / (raw_depth.max() - raw_depth.min()) * 255).astype("uint8")
     depth_o3d = o3d.geometry.Image(depth_image)
     image_o3d = o3d.geometry.Image(rgb_image)
+
+    # Create RGBD 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)
+    height, width = depth_image.shape
+
+    # Define camera intrinsics
+    camera_intrinsic = o3d.camera.PinholeCameraIntrinsic(
+        width,
+        height,
+        fx=1.0,
+        fy=1.0,
+        cx=width / 2.0,
+        cy=height / 2.0,
+    )
 
+    # Generate point cloud from RGBD image
     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
+    # Scale the Z dimension
+    points = np.asarray(pcd.points)
+    depth_scaled = ((raw_depth - raw_depth.min()) / (raw_depth.max() - raw_depth.min())) * z_scale
+    z_values = depth_scaled.flatten()[:len(points)]
+    points[:, 2] *= z_values
+    pcd.points = o3d.utility.Vector3dVector(points)
+
+    # Estimate and orient 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, 0.0, 1000.0])
+    pcd.orient_normals_towards_camera_location(camera_location=np.array([0.0, 0.0, 2.0 ]))
+
+    # Apply transformations
+    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]])
+
+    # Perform Poisson surface reconstruction
+    print(f"Running Poisson surface reconstruction with depth {depth}")
+    mesh_raw, densities = o3d.geometry.TriangleMesh.create_from_point_cloud_poisson(
+        pcd, depth=depth, width=0, scale=1.1, linear_fit=True
     )
-    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("run Poisson surface reconstruction")
-    with o3d.utility.VerbosityContextManager(o3d.utility.VerbosityLevel.Debug) as cm:
-        mesh_raw, densities = o3d.geometry.TriangleMesh.create_from_point_cloud_poisson(
-            pcd, depth=depth, width=0, scale=1.1, linear_fit=True
-        )
+    print(f"Raw mesh vertices: {len(mesh_raw.vertices)}, triangles: {len(mesh_raw.triangles)}")
 
-    voxel_size = max(mesh_raw.get_max_bound() - mesh_raw.get_min_bound()) / 256
-    print(f"voxel_size = {voxel_size:e}")
+    # Simplify the mesh using vertex clustering
+    voxel_size = max(mesh_raw.get_max_bound() - mesh_raw.get_min_bound()) / (max(width, height) * 0.8)
     mesh = mesh_raw.simplify_vertex_clustering(
         voxel_size=voxel_size,
         contraction=o3d.geometry.SimplificationContraction.Average,
     )
+    print(f"Simplified mesh vertices: {len(mesh.vertices)}, triangles: {len(mesh.triangles)}")
 
-    # vertices_to_remove = densities < np.quantile(densities, 0.001)
-    # mesh.remove_vertices_by_mask(vertices_to_remove)
+    # Crop the mesh to the bounding box of the point cloud
     bbox = pcd.get_axis_aligned_bounding_box()
     mesh_crop = mesh.crop(bbox)
-    gltf_path = f"./{image_path.stem}.gltf"
+
+    # Save the mesh as a GLTF file
+    gltf_path = f"./models/{image_path.stem}.gltf"
     o3d.io.write_triangle_mesh(gltf_path, mesh_crop, write_triangle_uvs=True)
     return gltf_path
 
 
-title = "Demo: zero-shot depth estimation with DPT + 3D Point Cloud"
-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 uses 3D Point Cloud to create a 3D object."
+# Define Gradio interface components
+title = "Demo: Zero-Shot Depth Estimation with DPT + 3D Point Cloud"
+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 uses 3D Point Cloud to create a 3D object."
+)
+# Create Gradio sliders for resized_width and z_scale
+resized_width_slider = gr.Slider(
+    minimum=400,
+    maximum=1600,
+    step=16,
+    value=800,
+    label="Resized Width",
+    info="Adjust the width to which the input image is resized."
+)
+
+z_scale_slider = gr.Slider(
+    minimum=160,
+    maximum=1024,
+    step=16,
+    value=208,
+    label="Z-Scale",
+    info="Adjust the scaling factor for the Z-axis in the 3D model."
+)
 examples = [["examples/" + img] for img in os.listdir("examples/")]
 
 iface = gr.Interface(
     fn=process_image,
-    inputs=[gr.Image(type="filepath", label="Input Image")],
+        inputs=[
+        gr.Image(type="filepath", label="Input Image"),
+        resized_width_slider,
+        z_scale_slider
+    ],
     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"),
+        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,
+    theme="Surn/Beeuty"
 )
-iface.launch(debug=True, show_api=False)
+
+if __name__ == "__main__":
+    iface.launch(debug=True, show_api=False, favicon_path="./favicon.ico")

requirements.txt

@@ -2,6 +2,6 @@ torch
 transformers
 numpy
 Pillow
-gradio==4.26.0
+gradio>=5.16.0
 jinja2
 open3d

web-ui.bat

@@ -0,0 +1,2 @@
+python311 -m app.py
+pause