Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -4,7 +4,6 @@ import numpy as np
|
|
4 |
from transformers import DPTForDepthEstimation, DPTImageProcessor
|
5 |
import gradio as gr
|
6 |
import torch.nn.utils.prune as prune
|
7 |
-
import open3d as o3d
|
8 |
|
9 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
10 |
|
@@ -32,31 +31,11 @@ model = model.to(device)
|
|
32 |
|
33 |
processor = DPTImageProcessor.from_pretrained("Intel/dpt-swinv2-tiny-256")
|
34 |
|
35 |
-
color_map = cv2.applyColorMap(np.arange(256, dtype=np.uint8), cv2.COLORMAP_INFERNO)
|
36 |
-
color_map = torch.from_numpy(color_map).to(device)
|
37 |
-
|
38 |
def preprocess_image(image):
|
39 |
image = cv2.resize(image, (128, 128))
|
40 |
image = torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0).float().to(device)
|
41 |
return image / 255.0
|
42 |
|
43 |
-
def create_point_cloud(depth_map, color_image):
|
44 |
-
rows, cols = depth_map.shape
|
45 |
-
c, r = np.meshgrid(np.arange(cols), np.arange(rows), sparse=True)
|
46 |
-
valid = (depth_map > 0) & (depth_map < 1000)
|
47 |
-
z = np.where(valid, depth_map, 0)
|
48 |
-
x = np.where(valid, z * (c - cols / 2) / cols, 0)
|
49 |
-
y = np.where(valid, z * (r - rows / 2) / rows, 0)
|
50 |
-
|
51 |
-
points = np.dstack((x, y, z)).reshape(-1, 3)
|
52 |
-
colors = color_image.reshape(-1, 3)
|
53 |
-
|
54 |
-
pcd = o3d.geometry.PointCloud()
|
55 |
-
pcd.points = o3d.utility.Vector3dVector(points)
|
56 |
-
pcd.colors = o3d.utility.Vector3dVector(colors / 255.0)
|
57 |
-
|
58 |
-
return pcd
|
59 |
-
|
60 |
@torch.inference_mode()
|
61 |
def process_frame(image):
|
62 |
if image is None:
|
@@ -68,24 +47,14 @@ def process_frame(image):
|
|
68 |
# Normalize depth map
|
69 |
depth_map = (depth_map - depth_map.min()) / (depth_map.max() - depth_map.min())
|
70 |
|
71 |
-
# Create
|
72 |
-
|
73 |
-
|
74 |
-
# Visualize point cloud
|
75 |
-
vis = o3d.visualization.Visualizer()
|
76 |
-
vis.create_window()
|
77 |
-
vis.add_geometry(pcd)
|
78 |
-
vis.poll_events()
|
79 |
-
vis.update_renderer()
|
80 |
-
|
81 |
-
# Capture the visualization as an image
|
82 |
-
image = vis.capture_screen_float_buffer(False)
|
83 |
-
vis.destroy_window()
|
84 |
|
85 |
-
#
|
86 |
-
|
|
|
87 |
|
88 |
-
return
|
89 |
|
90 |
interface = gr.Interface(
|
91 |
fn=process_frame,
|
@@ -94,4 +63,4 @@ interface = gr.Interface(
|
|
94 |
live=True
|
95 |
)
|
96 |
|
97 |
-
interface.launch()
|
|
|
4 |
from transformers import DPTForDepthEstimation, DPTImageProcessor
|
5 |
import gradio as gr
|
6 |
import torch.nn.utils.prune as prune
|
|
|
7 |
|
8 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
9 |
|
|
|
31 |
|
32 |
processor = DPTImageProcessor.from_pretrained("Intel/dpt-swinv2-tiny-256")
|
33 |
|
|
|
|
|
|
|
34 |
def preprocess_image(image):
|
35 |
image = cv2.resize(image, (128, 128))
|
36 |
image = torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0).float().to(device)
|
37 |
return image / 255.0
|
38 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
@torch.inference_mode()
|
40 |
def process_frame(image):
|
41 |
if image is None:
|
|
|
47 |
# Normalize depth map
|
48 |
depth_map = (depth_map - depth_map.min()) / (depth_map.max() - depth_map.min())
|
49 |
|
50 |
+
# Create a more visually informative depth map
|
51 |
+
depth_color = cv2.applyColorMap((depth_map * 255).astype(np.uint8), cv2.COLORMAP_INFERNO)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
|
53 |
+
# Blend original image with depth map for context
|
54 |
+
original_resized = cv2.resize(image, (128, 128))
|
55 |
+
blended = cv2.addWeighted(original_resized, 0.6, depth_color, 0.4, 0)
|
56 |
|
57 |
+
return blended
|
58 |
|
59 |
interface = gr.Interface(
|
60 |
fn=process_frame,
|
|
|
63 |
live=True
|
64 |
)
|
65 |
|
66 |
+
interface.launch()
|