Update app.py
Browse files
app.py
CHANGED
@@ -8,13 +8,14 @@ import os
|
|
8 |
import gradio as gr
|
9 |
import time
|
10 |
import io
|
|
|
11 |
|
12 |
# Set up device for torch
|
13 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
14 |
print(f"[INFO] Using device: {device}")
|
15 |
|
16 |
# Try to load the RAFT model from torch.hub.
|
17 |
-
# If it fails, fall back to OpenCV
|
18 |
try:
|
19 |
print("[INFO] Attempting to load RAFT model from torch.hub...")
|
20 |
raft_model = torch.hub.load("princeton-vl/RAFT", "raft_small", pretrained=True, trust_repo=True)
|
@@ -26,70 +27,48 @@ except Exception as e:
|
|
26 |
print("[INFO] Falling back to OpenCV Farneback optical flow.")
|
27 |
raft_model = None
|
28 |
|
29 |
-
def
|
30 |
"""
|
31 |
-
|
32 |
-
|
33 |
-
- Stabilizes the video using the generated motion data.
|
34 |
-
|
35 |
-
Yields:
|
36 |
-
A tuple of (original_video, stabilized_video, logs, progress)
|
37 |
-
During processing, original_video and stabilized_video are None.
|
38 |
-
The final yield returns the video file paths along with final logs and progress=100.
|
39 |
"""
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
# Check and extract the file path
|
46 |
-
if isinstance(video_file, dict):
|
47 |
-
video_file = video_file.get("name", None)
|
48 |
-
if video_file is None:
|
49 |
-
yield (None, None, "[ERROR] Please upload a video file.", 0)
|
50 |
-
return
|
51 |
-
|
52 |
-
add_log("[INFO] Starting AI-powered video processing...")
|
53 |
-
yield (None, None, add_log("Starting processing..."), 0)
|
54 |
-
|
55 |
-
# === CSV Generation Phase ===
|
56 |
-
add_log("[INFO] Starting motion CSV generation...")
|
57 |
-
yield (None, None, add_log("Starting CSV generation..."), 0)
|
58 |
|
59 |
cap = cv2.VideoCapture(video_file)
|
60 |
if not cap.isOpened():
|
61 |
-
|
62 |
-
return
|
63 |
-
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
64 |
-
add_log(f"[INFO] Total frames in video: {total_frames}")
|
65 |
|
66 |
-
|
67 |
-
|
68 |
-
with open(csv_file, 'w', newline='') as csvfile:
|
69 |
fieldnames = ['frame', 'mag', 'ang', 'zoom']
|
70 |
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
|
71 |
writer.writeheader()
|
72 |
|
73 |
ret, first_frame = cap.read()
|
74 |
if not ret:
|
75 |
-
|
76 |
-
return
|
77 |
|
78 |
if raft_model is not None:
|
79 |
first_frame_rgb = cv2.cvtColor(first_frame, cv2.COLOR_BGR2RGB)
|
80 |
prev_tensor = torch.from_numpy(first_frame_rgb).permute(2, 0, 1).float().unsqueeze(0) / 255.0
|
81 |
prev_tensor = prev_tensor.to(device)
|
82 |
-
|
83 |
else:
|
84 |
prev_gray = cv2.cvtColor(first_frame, cv2.COLOR_BGR2GRAY)
|
85 |
-
|
86 |
|
87 |
frame_idx = 1
|
88 |
-
|
|
|
89 |
while True:
|
90 |
ret, frame = cap.read()
|
91 |
if not ret:
|
92 |
break
|
|
|
93 |
if raft_model is not None:
|
94 |
curr_frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
95 |
curr_tensor = torch.from_numpy(curr_frame_rgb).permute(2, 0, 1).float().unsqueeze(0) / 255.0
|
@@ -104,12 +83,13 @@ def process_video_ai(video_file, zoom):
|
|
104 |
pyr_scale=0.5, levels=3, winsize=15,
|
105 |
iterations=3, poly_n=5, poly_sigma=1.2, flags=0)
|
106 |
prev_gray = curr_gray
|
107 |
-
|
108 |
-
# Compute median magnitude and angle
|
109 |
-
mag, ang = cv2.cartToPolar(flow[...,0], flow[...,1], angleInDegrees=True)
|
110 |
median_mag = np.median(mag)
|
111 |
median_ang = np.median(ang)
|
112 |
-
|
|
|
113 |
h, w = flow.shape[:2]
|
114 |
center_x, center_y = w / 2, h / 2
|
115 |
x_coords, y_coords = np.meshgrid(np.arange(w), np.arange(h))
|
@@ -126,23 +106,27 @@ def process_video_ai(video_file, zoom):
|
|
126 |
})
|
127 |
|
128 |
if frame_idx % 10 == 0 or frame_idx == total_frames:
|
129 |
-
|
130 |
-
add_log(f"[INFO] CSV: Processed frame {frame_idx}/{total_frames}")
|
131 |
-
yield (None, None, add_log(""), progress_csv)
|
132 |
frame_idx += 1
|
|
|
133 |
cap.release()
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
|
|
|
|
140 |
|
141 |
-
|
|
|
|
|
|
|
142 |
motion_data = {}
|
143 |
cumulative_dx = 0.0
|
144 |
cumulative_dy = 0.0
|
145 |
-
with open(
|
146 |
reader = csv.DictReader(csvfile)
|
147 |
for row in reader:
|
148 |
frame_num = int(row['frame'])
|
@@ -154,26 +138,43 @@ def process_video_ai(video_file, zoom):
|
|
154 |
cumulative_dx += dx
|
155 |
cumulative_dy += dy
|
156 |
motion_data[frame_num] = (-cumulative_dx, -cumulative_dy)
|
157 |
-
|
158 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
159 |
|
160 |
-
# Re-open video for stabilization
|
161 |
cap = cv2.VideoCapture(video_file)
|
|
|
|
|
|
|
162 |
fps = cap.get(cv2.CAP_PROP_FPS)
|
163 |
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
164 |
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
165 |
-
|
166 |
-
|
167 |
-
|
|
|
|
|
|
|
|
|
168 |
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
169 |
out = cv2.VideoWriter(output_file, fourcc, fps, (width, height))
|
170 |
|
171 |
frame_idx = 1
|
172 |
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
|
|
173 |
while True:
|
174 |
ret, frame = cap.read()
|
175 |
if not ret:
|
176 |
break
|
|
|
|
|
177 |
if zoom != 1.0:
|
178 |
zoomed_frame = cv2.resize(frame, None, fx=zoom, fy=zoom, interpolation=cv2.INTER_LINEAR)
|
179 |
zoomed_h, zoomed_w = zoomed_frame.shape[:2]
|
@@ -182,24 +183,48 @@ def process_video_ai(video_file, zoom):
|
|
182 |
frame = zoomed_frame[start_y:start_y+height, start_x:start_x+width]
|
183 |
|
184 |
dx, dy = motion_data.get(frame_idx, (0, 0))
|
185 |
-
transform = np.array([[1, 0, dx],
|
|
|
186 |
stabilized_frame = cv2.warpAffine(frame, transform, (width, height))
|
187 |
-
out.write(stabilized_frame)
|
188 |
|
|
|
189 |
if frame_idx % 10 == 0 or frame_idx == total_frames:
|
190 |
-
|
191 |
-
add_log(f"[INFO] Stabilization: Processed frame {frame_idx}/{total_frames}")
|
192 |
-
yield (None, None, add_log(""), progress_stab)
|
193 |
frame_idx += 1
|
|
|
194 |
cap.release()
|
195 |
out.release()
|
196 |
-
|
197 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
198 |
|
199 |
-
# Build the Gradio UI
|
200 |
with gr.Blocks() as demo:
|
201 |
gr.Markdown("# AI-Powered Video Stabilization")
|
202 |
-
gr.Markdown("Upload a video and select a zoom factor. The system will generate motion data using an AI model (RAFT if available
|
203 |
|
204 |
with gr.Row():
|
205 |
with gr.Column():
|
@@ -209,16 +234,12 @@ with gr.Blocks() as demo:
|
|
209 |
with gr.Column():
|
210 |
original_video = gr.Video(label="Original Video")
|
211 |
stabilized_video = gr.Video(label="Stabilized Video")
|
212 |
-
logs_output = gr.Textbox(label="Logs", lines=
|
213 |
-
progress_bar = gr.Slider(label="Progress", minimum=0, maximum=100, value=0, interactive=False)
|
214 |
-
|
215 |
-
demo.queue() # enable streaming
|
216 |
|
217 |
process_button.click(
|
218 |
fn=process_video_ai,
|
219 |
inputs=[video_input, zoom_slider],
|
220 |
-
outputs=[original_video, stabilized_video, logs_output
|
221 |
-
stream=True # enable streaming updates
|
222 |
)
|
223 |
|
224 |
-
demo.launch()
|
|
|
8 |
import gradio as gr
|
9 |
import time
|
10 |
import io
|
11 |
+
from contextlib import redirect_stdout
|
12 |
|
13 |
# Set up device for torch
|
14 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
15 |
print(f"[INFO] Using device: {device}")
|
16 |
|
17 |
# Try to load the RAFT model from torch.hub.
|
18 |
+
# If it fails, we fall back to OpenCV optical flow.
|
19 |
try:
|
20 |
print("[INFO] Attempting to load RAFT model from torch.hub...")
|
21 |
raft_model = torch.hub.load("princeton-vl/RAFT", "raft_small", pretrained=True, trust_repo=True)
|
|
|
27 |
print("[INFO] Falling back to OpenCV Farneback optical flow.")
|
28 |
raft_model = None
|
29 |
|
30 |
+
def generate_motion_csv(video_file, output_csv=None):
|
31 |
"""
|
32 |
+
Generates a CSV file with motion data (columns: frame, mag, ang, zoom) from an input video.
|
33 |
+
Uses RAFT if available, otherwise falls back to OpenCV's Farneback optical flow.
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
"""
|
35 |
+
start_time = time.time()
|
36 |
+
if output_csv is None:
|
37 |
+
temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.csv')
|
38 |
+
output_csv = temp_file.name
|
39 |
+
temp_file.close()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
cap = cv2.VideoCapture(video_file)
|
42 |
if not cap.isOpened():
|
43 |
+
raise ValueError("[ERROR] Could not open video file for CSV generation.")
|
|
|
|
|
|
|
44 |
|
45 |
+
print(f"[INFO] Generating motion CSV for video: {video_file}")
|
46 |
+
with open(output_csv, 'w', newline='') as csvfile:
|
|
|
47 |
fieldnames = ['frame', 'mag', 'ang', 'zoom']
|
48 |
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
|
49 |
writer.writeheader()
|
50 |
|
51 |
ret, first_frame = cap.read()
|
52 |
if not ret:
|
53 |
+
raise ValueError("[ERROR] Cannot read first frame from video.")
|
|
|
54 |
|
55 |
if raft_model is not None:
|
56 |
first_frame_rgb = cv2.cvtColor(first_frame, cv2.COLOR_BGR2RGB)
|
57 |
prev_tensor = torch.from_numpy(first_frame_rgb).permute(2, 0, 1).float().unsqueeze(0) / 255.0
|
58 |
prev_tensor = prev_tensor.to(device)
|
59 |
+
print("[INFO] Using RAFT model for optical flow computation.")
|
60 |
else:
|
61 |
prev_gray = cv2.cvtColor(first_frame, cv2.COLOR_BGR2GRAY)
|
62 |
+
print("[INFO] Using OpenCV Farneback optical flow for computation.")
|
63 |
|
64 |
frame_idx = 1
|
65 |
+
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
66 |
+
print(f"[INFO] Total frames to process: {total_frames}")
|
67 |
while True:
|
68 |
ret, frame = cap.read()
|
69 |
if not ret:
|
70 |
break
|
71 |
+
|
72 |
if raft_model is not None:
|
73 |
curr_frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
74 |
curr_tensor = torch.from_numpy(curr_frame_rgb).permute(2, 0, 1).float().unsqueeze(0) / 255.0
|
|
|
83 |
pyr_scale=0.5, levels=3, winsize=15,
|
84 |
iterations=3, poly_n=5, poly_sigma=1.2, flags=0)
|
85 |
prev_gray = curr_gray
|
86 |
+
|
87 |
+
# Compute median magnitude and angle of the optical flow.
|
88 |
+
mag, ang = cv2.cartToPolar(flow[..., 0], flow[..., 1], angleInDegrees=True)
|
89 |
median_mag = np.median(mag)
|
90 |
median_ang = np.median(ang)
|
91 |
+
|
92 |
+
# Compute a "zoom factor": fraction of pixels moving away from the center.
|
93 |
h, w = flow.shape[:2]
|
94 |
center_x, center_y = w / 2, h / 2
|
95 |
x_coords, y_coords = np.meshgrid(np.arange(w), np.arange(h))
|
|
|
106 |
})
|
107 |
|
108 |
if frame_idx % 10 == 0 or frame_idx == total_frames:
|
109 |
+
print(f"[INFO] Processed frame {frame_idx}/{total_frames}")
|
|
|
|
|
110 |
frame_idx += 1
|
111 |
+
|
112 |
cap.release()
|
113 |
+
elapsed = time.time() - start_time
|
114 |
+
print(f"[INFO] Motion CSV generated: {output_csv} in {elapsed:.2f} seconds")
|
115 |
+
return output_csv
|
116 |
+
|
117 |
+
def read_motion_csv(csv_filename):
|
118 |
+
"""
|
119 |
+
Reads a motion CSV file (with columns: frame, mag, ang, zoom) and computes a cumulative
|
120 |
+
offset per frame for stabilization.
|
121 |
|
122 |
+
Returns:
|
123 |
+
A dictionary mapping frame numbers to (dx, dy) offsets.
|
124 |
+
"""
|
125 |
+
print(f"[INFO] Reading motion CSV: {csv_filename}")
|
126 |
motion_data = {}
|
127 |
cumulative_dx = 0.0
|
128 |
cumulative_dy = 0.0
|
129 |
+
with open(csv_filename, 'r') as csvfile:
|
130 |
reader = csv.DictReader(csvfile)
|
131 |
for row in reader:
|
132 |
frame_num = int(row['frame'])
|
|
|
138 |
cumulative_dx += dx
|
139 |
cumulative_dy += dy
|
140 |
motion_data[frame_num] = (-cumulative_dx, -cumulative_dy)
|
141 |
+
print("[INFO] Completed reading motion CSV.")
|
142 |
+
return motion_data
|
143 |
+
|
144 |
+
def stabilize_video_using_csv(video_file, csv_file, zoom=1.0, output_file=None):
|
145 |
+
"""
|
146 |
+
Stabilizes the input video using motion data from the CSV file.
|
147 |
+
"""
|
148 |
+
start_time = time.time()
|
149 |
+
print(f"[INFO] Starting stabilization using CSV: {csv_file}")
|
150 |
+
motion_data = read_motion_csv(csv_file)
|
151 |
|
|
|
152 |
cap = cv2.VideoCapture(video_file)
|
153 |
+
if not cap.isOpened():
|
154 |
+
raise ValueError("[ERROR] Could not open video file for stabilization.")
|
155 |
+
|
156 |
fps = cap.get(cv2.CAP_PROP_FPS)
|
157 |
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
158 |
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
159 |
+
print(f"[INFO] Video properties - FPS: {fps}, Width: {width}, Height: {height}")
|
160 |
+
|
161 |
+
if output_file is None:
|
162 |
+
temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.mp4')
|
163 |
+
output_file = temp_file.name
|
164 |
+
temp_file.close()
|
165 |
+
|
166 |
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
167 |
out = cv2.VideoWriter(output_file, fourcc, fps, (width, height))
|
168 |
|
169 |
frame_idx = 1
|
170 |
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
171 |
+
print(f"[INFO] Total frames to stabilize: {total_frames}")
|
172 |
while True:
|
173 |
ret, frame = cap.read()
|
174 |
if not ret:
|
175 |
break
|
176 |
+
|
177 |
+
# Optionally apply zoom (resize and center-crop)
|
178 |
if zoom != 1.0:
|
179 |
zoomed_frame = cv2.resize(frame, None, fx=zoom, fy=zoom, interpolation=cv2.INTER_LINEAR)
|
180 |
zoomed_h, zoomed_w = zoomed_frame.shape[:2]
|
|
|
183 |
frame = zoomed_frame[start_y:start_y+height, start_x:start_x+width]
|
184 |
|
185 |
dx, dy = motion_data.get(frame_idx, (0, 0))
|
186 |
+
transform = np.array([[1, 0, dx],
|
187 |
+
[0, 1, dy]], dtype=np.float32)
|
188 |
stabilized_frame = cv2.warpAffine(frame, transform, (width, height))
|
|
|
189 |
|
190 |
+
out.write(stabilized_frame)
|
191 |
if frame_idx % 10 == 0 or frame_idx == total_frames:
|
192 |
+
print(f"[INFO] Stabilized frame {frame_idx}/{total_frames}")
|
|
|
|
|
193 |
frame_idx += 1
|
194 |
+
|
195 |
cap.release()
|
196 |
out.release()
|
197 |
+
elapsed = time.time() - start_time
|
198 |
+
print(f"[INFO] Stabilized video saved to: {output_file} in {elapsed:.2f} seconds")
|
199 |
+
return output_file
|
200 |
+
|
201 |
+
def process_video_ai(video_file, zoom):
|
202 |
+
"""
|
203 |
+
Gradio interface function:
|
204 |
+
- Generates motion data (CSV) from the input video using an AI model (RAFT if available, else Farneback).
|
205 |
+
- Stabilizes the video based on the generated motion data.
|
206 |
+
|
207 |
+
Returns:
|
208 |
+
Tuple containing the original video file path, the stabilized video file path, and log output.
|
209 |
+
"""
|
210 |
+
log_buffer = io.StringIO()
|
211 |
+
with redirect_stdout(log_buffer):
|
212 |
+
if isinstance(video_file, dict):
|
213 |
+
video_file = video_file.get("name", None)
|
214 |
+
if video_file is None:
|
215 |
+
raise ValueError("[ERROR] Please upload a video file.")
|
216 |
+
|
217 |
+
print("[INFO] Starting AI-powered video processing...")
|
218 |
+
csv_file = generate_motion_csv(video_file)
|
219 |
+
stabilized_path = stabilize_video_using_csv(video_file, csv_file, zoom=zoom)
|
220 |
+
print("[INFO] Video processing complete.")
|
221 |
+
logs = log_buffer.getvalue()
|
222 |
+
return video_file, stabilized_path, logs
|
223 |
|
224 |
+
# Build the Gradio UI.
|
225 |
with gr.Blocks() as demo:
|
226 |
gr.Markdown("# AI-Powered Video Stabilization")
|
227 |
+
gr.Markdown("Upload a video and select a zoom factor. The system will generate motion data using an AI model (RAFT if available) and then stabilize the video.")
|
228 |
|
229 |
with gr.Row():
|
230 |
with gr.Column():
|
|
|
234 |
with gr.Column():
|
235 |
original_video = gr.Video(label="Original Video")
|
236 |
stabilized_video = gr.Video(label="Stabilized Video")
|
237 |
+
logs_output = gr.Textbox(label="Logs", lines=10)
|
|
|
|
|
|
|
238 |
|
239 |
process_button.click(
|
240 |
fn=process_video_ai,
|
241 |
inputs=[video_input, zoom_slider],
|
242 |
+
outputs=[original_video, stabilized_video, logs_output]
|
|
|
243 |
)
|
244 |
|
245 |
+
demo.launch()
|