phrographica commited on
Commit
fca5a33
·
verified ·
1 Parent(s): 59e15d1

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +121 -13
app.py CHANGED
@@ -14,6 +14,9 @@
14
  import spaces
15
  import gradio as gr
16
  import gc
 
 
 
17
 
18
  import numpy as np
19
  import os
@@ -100,26 +103,119 @@ def infer_video_depth(
100
  max_len: int = -1,
101
  target_fps: int = -1,
102
  max_res: int = 1280,
103
- grayscale: bool = False,
104
  output_dir: str = './outputs',
105
  input_size: int = 518,
106
  ):
107
- frames, target_fps = read_video_frames(input_video, max_len, target_fps, max_res)
108
- depths, fps = video_depth_anything.infer_video_depth(frames, target_fps, input_size=input_size, device=DEVICE)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
109
 
110
  video_name = os.path.basename(input_video)
111
- if not os.path.exists(output_dir):
112
- os.makedirs(output_dir)
 
 
 
 
 
 
 
 
 
 
 
113
 
114
- processed_video_path = os.path.join(output_dir, os.path.splitext(video_name)[0]+'_src.mp4')
115
- depth_vis_path = os.path.join(output_dir, os.path.splitext(video_name)[0]+'_vis.mp4')
116
  save_video(frames, processed_video_path, fps=fps)
117
- save_video(depths, depth_vis_path, fps=fps, is_depths=True, grayscale=grayscale)
 
 
 
 
 
 
 
 
 
118
 
119
- gc.collect()
120
- torch.cuda.empty_cache()
121
 
122
- return [processed_video_path, depth_vis_path]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
123
 
124
 
125
  def construct_demo():
@@ -151,6 +247,10 @@ def construct_demo():
151
  show_share_button=True,
152
  scale=5,
153
  )
 
 
 
 
154
 
155
  with gr.Row(equal_height=True):
156
  with gr.Column(scale=1):
@@ -193,7 +293,11 @@ def construct_demo():
193
  target_fps,
194
  max_res
195
  ],
196
- outputs=[processed_video, depth_vis_video],
 
 
 
 
197
  fn=infer_video_depth,
198
  cache_examples="lazy",
199
  )
@@ -207,7 +311,11 @@ def construct_demo():
207
  max_res,
208
  grayscale
209
  ],
210
- outputs=[processed_video, depth_vis_video],
 
 
 
 
211
  )
212
 
213
  return demo
 
14
  import spaces
15
  import gradio as gr
16
  import gc
17
+ import zipfile
18
+ import OpenEXR
19
+ import Imath
20
 
21
  import numpy as np
22
  import os
 
103
  max_len: int = -1,
104
  target_fps: int = -1,
105
  max_res: int = 1280,
 
106
  output_dir: str = './outputs',
107
  input_size: int = 518,
108
  ):
109
+ frames, target_fps = read_video_frames(
110
+ input_video,
111
+ max_len,
112
+ target_fps,
113
+ max_res
114
+ )
115
+
116
+ depth_list, fps = video_depth_anything.infer_video_depth(
117
+ frames,
118
+ target_fps,
119
+ input_size=input_size,
120
+ device=DEVICE
121
+ )
122
+
123
+ # Raw model output
124
+ depth_list = np.stack(depth_list, axis=0).astype(np.float32)
125
+
126
+ # MP4 visualization
127
+ vis = vis_sequence_depth(depth_list)
128
 
129
  video_name = os.path.basename(input_video)
130
+ base_name = os.path.splitext(video_name)[0]
131
+
132
+ os.makedirs(output_dir, exist_ok=True)
133
+
134
+ processed_video_path = os.path.join(
135
+ output_dir,
136
+ base_name + '_src.mp4'
137
+ )
138
+
139
+ depth_vis_path = os.path.join(
140
+ output_dir,
141
+ base_name + '_vis.mp4'
142
+ )
143
 
 
 
144
  save_video(frames, processed_video_path, fps=fps)
145
+ save_video(vis, depth_vis_path, fps=fps)
146
+
147
+ # ----------------------------------------------------
148
+ # RAW 32-BIT FLOAT EXR EXPORT
149
+ # ----------------------------------------------------
150
+
151
+ depth_exr_dir = os.path.join(
152
+ output_dir,
153
+ base_name + '_depth_EXR'
154
+ )
155
 
156
+ os.makedirs(depth_exr_dir, exist_ok=True)
 
157
 
158
+ FLOAT = Imath.PixelType(Imath.PixelType.FLOAT)
159
+
160
+ for i, depth in enumerate(depth_list):
161
+
162
+ depth = np.ascontiguousarray(depth.astype(np.float32))
163
+
164
+ height, width = depth.shape
165
+
166
+ exr_path = os.path.join(
167
+ depth_exr_dir,
168
+ f"frame_{i:05d}.exr"
169
+ )
170
+
171
+ header = OpenEXR.Header(width, height)
172
+
173
+ # Store depth in RGB so Blender can easily read it
174
+ # through an Image Texture node.
175
+ header["channels"] = {
176
+ "R": Imath.Channel(FLOAT),
177
+ "G": Imath.Channel(FLOAT),
178
+ "B": Imath.Channel(FLOAT),
179
+ }
180
+
181
+ depth_bytes = depth.tobytes()
182
+
183
+ exr = OpenEXR.OutputFile(exr_path, header)
184
+
185
+ exr.writePixels({
186
+ "R": depth_bytes,
187
+ "G": depth_bytes,
188
+ "B": depth_bytes,
189
+ })
190
+
191
+ exr.close()
192
+
193
+ # ZIP the EXRs because Gradio cannot conveniently
194
+ # return a directory as one downloadable result.
195
+ zip_path = os.path.join(
196
+ output_dir,
197
+ base_name + '_depth_EXR.zip'
198
+ )
199
+
200
+ with zipfile.ZipFile(
201
+ zip_path,
202
+ 'w',
203
+ compression=zipfile.ZIP_DEFLATED
204
+ ) as zipf:
205
+
206
+ for filename in sorted(os.listdir(depth_exr_dir)):
207
+ file_path = os.path.join(depth_exr_dir, filename)
208
+
209
+ zipf.write(
210
+ file_path,
211
+ arcname=filename
212
+ )
213
+
214
+ return [
215
+ processed_video_path,
216
+ depth_vis_path,
217
+ zip_path
218
+ ]
219
 
220
 
221
  def construct_demo():
 
247
  show_share_button=True,
248
  scale=5,
249
  )
250
+ depth_exr_file = gr.File(
251
+ label="Raw 32-bit EXR Depth Sequence",
252
+ interactive=False
253
+ )
254
 
255
  with gr.Row(equal_height=True):
256
  with gr.Column(scale=1):
 
293
  target_fps,
294
  max_res
295
  ],
296
+ outputs=[
297
+ processed_video,
298
+ depth_vis_video,
299
+ depth_exr_file
300
+ ],
301
  fn=infer_video_depth,
302
  cache_examples="lazy",
303
  )
 
311
  max_res,
312
  grayscale
313
  ],
314
+ outputs=[
315
+ processed_video,
316
+ depth_vis_video,
317
+ depth_exr_file
318
+ ],
319
  )
320
 
321
  return demo