Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import cv2
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
from sam2.sam2_video_predictor import SAM2VideoPredictor
|
6 |
+
from ultralytics import YOLO
|
7 |
+
import supervision as sv
|
8 |
+
import os
|
9 |
+
|
10 |
+
# Initialize models from Hugging Face Hub
|
11 |
+
predictor = SAM2VideoPredictor.from_pretrained("facebook/sam2.1-hiera-tiny")
|
12 |
+
yolo_model = YOLO("ultralytics/yolo-world-v8n") # Lightweight YOLO-World model
|
13 |
+
|
14 |
+
def detect_motorcycles(frame, prompt="motorcycle"):
|
15 |
+
"""Detect motorcycles in a frame using YOLO-World and return bounding boxes."""
|
16 |
+
results = yolo_model.predict(frame, prompt=prompt, device="cpu")
|
17 |
+
boxes = []
|
18 |
+
for result in results:
|
19 |
+
for box in result.boxes:
|
20 |
+
if result.names[box.cls] == prompt:
|
21 |
+
x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
|
22 |
+
boxes.append([x1, y1, x2, y2])
|
23 |
+
return boxes
|
24 |
+
|
25 |
+
def segment_and_highlight_video(video_path, prompt="motorcycle", highlight_color="red"):
|
26 |
+
"""Segment and highlight motorcycles in a video using SAM 2 and YOLO-World."""
|
27 |
+
# Create temporary directory for video frames
|
28 |
+
frames_dir = "video_frames"
|
29 |
+
os.makedirs(frames_dir, exist_ok=True)
|
30 |
+
|
31 |
+
# Extract frames
|
32 |
+
cap = cv2.VideoCapture(video_path)
|
33 |
+
fps = cap.get(cv2.CAP_PROP_FPS)
|
34 |
+
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
35 |
+
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
36 |
+
# Limit resolution for CPU
|
37 |
+
if width > 640:
|
38 |
+
height = int(height * 640 / width)
|
39 |
+
width = 640
|
40 |
+
frame_paths = []
|
41 |
+
|
42 |
+
# Save frames as JPEG
|
43 |
+
frame_idx = 0
|
44 |
+
with sv.ImageSink(target_dir_path=frames_dir, image_name_pattern="{:05d}.jpeg") as sink:
|
45 |
+
while cap.isOpened():
|
46 |
+
ret, frame = cap.read()
|
47 |
+
if not ret:
|
48 |
+
break
|
49 |
+
frame = cv2.resize(frame, (width, height))
|
50 |
+
sink.save_image(frame)
|
51 |
+
frame_paths.append(os.path.join(frames_dir, f"{frame_idx:05d}.jpeg"))
|
52 |
+
frame_idx += 1
|
53 |
+
cap.release()
|
54 |
+
|
55 |
+
# Initialize SAM 2 inference state
|
56 |
+
with torch.inference_mode():
|
57 |
+
state = predictor.init_state(video_path=frames_dir)
|
58 |
+
|
59 |
+
# Detect motorcycles in the first frame
|
60 |
+
first_frame = cv2.imread(frame_paths[0])
|
61 |
+
boxes = detect_motorcycles(first_frame, prompt)
|
62 |
+
|
63 |
+
# Add boxes as prompts for SAM 2
|
64 |
+
if boxes:
|
65 |
+
frame_idx, obj_ids, masks = predictor.add_new_points_or_box(
|
66 |
+
state, frame_idx=0, obj_ids=[1], boxes=np.array(boxes)
|
67 |
+
)
|
68 |
+
|
69 |
+
# Create output video
|
70 |
+
output_path = "output.mp4"
|
71 |
+
out = cv2.VideoWriter(output_path, cv2.VideoWriter_fourcc(*"mp4v"), fps, (width, height))
|
72 |
+
|
73 |
+
# Color map for highlighting
|
74 |
+
color_map = {"red": (0, 0, 255), "green": (0, 255, 0), "blue": (255, 0, 0)}
|
75 |
+
highlight_rgb = color_map.get(highlight_color.lower(), (0, 0, 255))
|
76 |
+
|
77 |
+
# Propagate masks and apply highlights
|
78 |
+
for frame_idx, obj_ids, masks in predictor.propagate_in_video(state):
|
79 |
+
frame = cv2.imread(frame_paths[frame_idx])
|
80 |
+
mask = masks[0].astype(np.uint8) * 255 # Assuming one object
|
81 |
+
mask_colored = np.zeros_like(frame)
|
82 |
+
mask_colored[:, :, 0] = mask * highlight_rgb[0]
|
83 |
+
mask_colored[:, :, 1] = mask * highlight_rgb[1]
|
84 |
+
mask_colored[:, :, 2] = mask * highlight_rgb[2]
|
85 |
+
highlighted_frame = cv2.addWeighted(frame, 0.7, mask_colored, 0.3, 0)
|
86 |
+
out.write(highlighted_frame)
|
87 |
+
|
88 |
+
out.release()
|
89 |
+
|
90 |
+
# Clean up
|
91 |
+
for frame_path in frame_paths:
|
92 |
+
os.remove(frame_path)
|
93 |
+
os.rmdir(frames_dir)
|
94 |
+
|
95 |
+
return output_path
|
96 |
+
|
97 |
+
# Gradio interface
|
98 |
+
iface = gr.Interface(
|
99 |
+
fn=segment_and_highlight_video,
|
100 |
+
inputs=[
|
101 |
+
gr.Video(label="Upload Video"),
|
102 |
+
gr.Textbox(label="Prompt", placeholder="e.g., motorcycle"),
|
103 |
+
gr.Dropdown(choices=["red", "green", "blue"], label="Highlight Color")
|
104 |
+
],
|
105 |
+
outputs=gr.Video(label="Highlighted Video"),
|
106 |
+
title="Video Segmentation with SAM 2 and YOLO-World (CPU)",
|
107 |
+
description="Upload a short video, specify a text prompt (e.g., 'motorcycle'), and choose a highlight color. Optimized for CPU."
|
108 |
+
)
|
109 |
+
iface.launch()
|