Spaces:
Running
on
L40S
Running
on
L40S
update video writer
Browse files- app.py +8 -6
- requirements.txt +2 -1
- utils/dc_utils.py +67 -65
- video_depth_anything/video_depth.py +1 -1
app.py
CHANGED
@@ -13,14 +13,14 @@
|
|
13 |
# limitations under the License.
|
14 |
import spaces
|
15 |
import gradio as gr
|
16 |
-
|
17 |
|
18 |
import numpy as np
|
19 |
import os
|
20 |
import torch
|
21 |
|
22 |
from video_depth_anything.video_depth import VideoDepthAnything
|
23 |
-
from utils.dc_utils import read_video_frames,
|
24 |
|
25 |
from huggingface_hub import hf_hub_download
|
26 |
|
@@ -73,9 +73,8 @@ def infer_video_depth(
|
|
73 |
input_size: int = 518,
|
74 |
):
|
75 |
frames, target_fps = read_video_frames(input_video, max_len, target_fps, max_res)
|
76 |
-
|
77 |
-
|
78 |
-
vis = vis_sequence_depth(depth_list)
|
79 |
video_name = os.path.basename(input_video)
|
80 |
if not os.path.exists(output_dir):
|
81 |
os.makedirs(output_dir)
|
@@ -83,7 +82,10 @@ def infer_video_depth(
|
|
83 |
processed_video_path = os.path.join(output_dir, os.path.splitext(video_name)[0]+'_src.mp4')
|
84 |
depth_vis_path = os.path.join(output_dir, os.path.splitext(video_name)[0]+'_vis.mp4')
|
85 |
save_video(frames, processed_video_path, fps=fps)
|
86 |
-
save_video(
|
|
|
|
|
|
|
87 |
|
88 |
return [processed_video_path, depth_vis_path]
|
89 |
|
|
|
13 |
# limitations under the License.
|
14 |
import spaces
|
15 |
import gradio as gr
|
16 |
+
import gc
|
17 |
|
18 |
import numpy as np
|
19 |
import os
|
20 |
import torch
|
21 |
|
22 |
from video_depth_anything.video_depth import VideoDepthAnything
|
23 |
+
from utils.dc_utils import read_video_frames, save_video
|
24 |
|
25 |
from huggingface_hub import hf_hub_download
|
26 |
|
|
|
73 |
input_size: int = 518,
|
74 |
):
|
75 |
frames, target_fps = read_video_frames(input_video, max_len, target_fps, max_res)
|
76 |
+
depths, fps = video_depth_anything.infer_video_depth(frames, target_fps, input_size=input_size, device=DEVICE)
|
77 |
+
|
|
|
78 |
video_name = os.path.basename(input_video)
|
79 |
if not os.path.exists(output_dir):
|
80 |
os.makedirs(output_dir)
|
|
|
82 |
processed_video_path = os.path.join(output_dir, os.path.splitext(video_name)[0]+'_src.mp4')
|
83 |
depth_vis_path = os.path.join(output_dir, os.path.splitext(video_name)[0]+'_vis.mp4')
|
84 |
save_video(frames, processed_video_path, fps=fps)
|
85 |
+
save_video(depths, depth_vis_path, fps=fps, is_depths=True)
|
86 |
+
|
87 |
+
gc.collect()
|
88 |
+
torch.cuda.empty_cache()
|
89 |
|
90 |
return [processed_video_path, depth_vis_path]
|
91 |
|
requirements.txt
CHANGED
@@ -7,7 +7,8 @@ opencv-python
|
|
7 |
matplotlib
|
8 |
huggingface_hub
|
9 |
pillow
|
10 |
-
|
|
|
11 |
decord
|
12 |
xformers
|
13 |
einops
|
|
|
7 |
matplotlib
|
8 |
huggingface_hub
|
9 |
pillow
|
10 |
+
imageio
|
11 |
+
imageio-ffmpeg
|
12 |
decord
|
13 |
xformers
|
14 |
einops
|
utils/dc_utils.py
CHANGED
@@ -3,82 +3,84 @@
|
|
3 |
#
|
4 |
# This file may have been modified by ByteDance Ltd. and/or its affiliates on [date of modification]
|
5 |
# Original file is released under [ MIT License license], with the full license text available at [https://github.com/Tencent/DepthCrafter?tab=License-1-ov-file].
|
6 |
-
from typing import Union, List
|
7 |
-
import tempfile
|
8 |
import numpy as np
|
9 |
-
import PIL.Image
|
10 |
import matplotlib.cm as cm
|
11 |
-
import
|
12 |
-
|
13 |
-
from decord import VideoReader, cpu
|
|
|
|
|
|
|
|
|
14 |
|
|
|
|
|
15 |
|
16 |
-
def read_video_frames(video_path, process_length, target_fps=-1, max_res=-1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
|
18 |
-
|
19 |
-
print("==> original video shape: ", (len(vid), *vid.get_batch([0]).shape[1:]))
|
20 |
-
original_height, original_width = vid.get_batch([0]).shape[1:3]
|
21 |
-
height = original_height
|
22 |
-
width = original_width
|
23 |
-
if max_res > 0 and max(height, width) > max_res:
|
24 |
-
scale = max_res / max(original_height, original_width)
|
25 |
-
height = round(original_height * scale)
|
26 |
-
width = round(original_width * scale)
|
27 |
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
if process_length != -1 and process_length < len(frames_idx):
|
35 |
-
frames_idx = frames_idx[:process_length]
|
36 |
-
print(f"==> final processing shape: {len(frames_idx), *vid.get_batch([0]).shape[1:]}")
|
37 |
-
frames = vid.get_batch(frames_idx).asnumpy()
|
38 |
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
def save_video(
|
43 |
-
video_frames: Union[List[np.ndarray], List[PIL.Image.Image]],
|
44 |
-
output_video_path: str = None,
|
45 |
-
fps: int = 10,
|
46 |
-
crf: int = 18,
|
47 |
-
) -> str:
|
48 |
-
if output_video_path is None:
|
49 |
-
output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name
|
50 |
|
51 |
-
|
52 |
-
video_frames = [frame.astype(np.uint8) for frame in video_frames]
|
53 |
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
|
|
|
59 |
|
60 |
-
class ColorMapper:
|
61 |
-
# a color mapper to map depth values to a certain colormap
|
62 |
-
def __init__(self, colormap: str = "inferno"):
|
63 |
-
self.colormap = torch.tensor(cm.get_cmap(colormap).colors)
|
64 |
-
|
65 |
-
def apply(self, image: torch.Tensor, v_min=None, v_max=None):
|
66 |
-
# assert len(image.shape) == 2
|
67 |
-
if v_min is None:
|
68 |
-
v_min = image.min()
|
69 |
-
if v_max is None:
|
70 |
-
v_max = image.max()
|
71 |
-
image = (image - v_min) / (v_max - v_min)
|
72 |
-
image = (image * 255).long()
|
73 |
-
image = self.colormap[image] * 255
|
74 |
-
return image
|
75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
-
|
78 |
-
visualizer = ColorMapper()
|
79 |
-
if v_min is None:
|
80 |
-
v_min = depths.min()
|
81 |
-
if v_max is None:
|
82 |
-
v_max = depths.max()
|
83 |
-
res = visualizer.apply(torch.tensor(depths), v_min=v_min, v_max=v_max).numpy()
|
84 |
-
return res
|
|
|
3 |
#
|
4 |
# This file may have been modified by ByteDance Ltd. and/or its affiliates on [date of modification]
|
5 |
# Original file is released under [ MIT License license], with the full license text available at [https://github.com/Tencent/DepthCrafter?tab=License-1-ov-file].
|
|
|
|
|
6 |
import numpy as np
|
|
|
7 |
import matplotlib.cm as cm
|
8 |
+
import imageio
|
9 |
+
try:
|
10 |
+
from decord import VideoReader, cpu
|
11 |
+
DECORD_AVAILABLE = True
|
12 |
+
except:
|
13 |
+
import cv2
|
14 |
+
DECORD_AVAILABLE = False
|
15 |
|
16 |
+
def ensure_even(value):
|
17 |
+
return value if value % 2 == 0 else value + 1
|
18 |
|
19 |
+
def read_video_frames(video_path, process_length, target_fps=-1, max_res=-1):
|
20 |
+
if DECORD_AVAILABLE:
|
21 |
+
vid = VideoReader(video_path, ctx=cpu(0))
|
22 |
+
original_height, original_width = vid.get_batch([0]).shape[1:3]
|
23 |
+
height = original_height
|
24 |
+
width = original_width
|
25 |
+
if max_res > 0 and max(height, width) > max_res:
|
26 |
+
scale = max_res / max(original_height, original_width)
|
27 |
+
height = ensure_even(round(original_height * scale))
|
28 |
+
width = ensure_even(round(original_width * scale))
|
29 |
|
30 |
+
vid = VideoReader(video_path, ctx=cpu(0), width=width, height=height)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
+
fps = vid.get_avg_fps() if target_fps == -1 else target_fps
|
33 |
+
stride = round(vid.get_avg_fps() / fps)
|
34 |
+
stride = max(stride, 1)
|
35 |
+
frames_idx = list(range(0, len(vid), stride))
|
36 |
+
if process_length != -1 and process_length < len(frames_idx):
|
37 |
+
frames_idx = frames_idx[:process_length]
|
38 |
+
frames = vid.get_batch(frames_idx).asnumpy()
|
39 |
+
else:
|
40 |
+
cap = cv2.VideoCapture(video_path)
|
41 |
+
original_fps = cap.get(cv2.CAP_PROP_FPS)
|
42 |
+
original_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
43 |
+
original_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
44 |
|
45 |
+
if max_res > 0 and max(original_height, original_width) > max_res:
|
46 |
+
scale = max_res / max(original_height, original_width)
|
47 |
+
height = round(original_height * scale)
|
48 |
+
width = round(original_width * scale)
|
|
|
|
|
|
|
|
|
49 |
|
50 |
+
fps = original_fps if target_fps < 0 else target_fps
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
|
52 |
+
stride = max(round(original_fps / fps), 1)
|
|
|
53 |
|
54 |
+
frames = []
|
55 |
+
frame_count = 0
|
56 |
+
while cap.isOpened():
|
57 |
+
ret, frame = cap.read()
|
58 |
+
if not ret or (process_length > 0 and frame_count >= process_length):
|
59 |
+
break
|
60 |
+
if frame_count % stride == 0:
|
61 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) # Convert BGR to RGB
|
62 |
+
if max_res > 0 and max(original_height, original_width) > max_res:
|
63 |
+
frame = cv2.resize(frame, (width, height)) # Resize frame
|
64 |
+
frames.append(frame)
|
65 |
+
frame_count += 1
|
66 |
+
cap.release()
|
67 |
+
frames = np.stack(frames, axis=0)
|
68 |
|
69 |
+
return frames, fps
|
70 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
|
72 |
+
def save_video(frames, output_video_path, fps=10, is_depths=False):
|
73 |
+
writer = imageio.get_writer(output_video_path, fps=fps, macro_block_size=1, codec='libx264', ffmpeg_params=['-crf', '18'])
|
74 |
+
if is_depths:
|
75 |
+
colormap = np.array(cm.get_cmap("inferno").colors)
|
76 |
+
d_min, d_max = frames.min(), frames.max()
|
77 |
+
for i in range(frames.shape[0]):
|
78 |
+
depth = frames[i]
|
79 |
+
depth_norm = ((depth - d_min) / (d_max - d_min) * 255).astype(np.uint8)
|
80 |
+
depth_vis = (colormap[depth_norm] * 255).astype(np.uint8)
|
81 |
+
writer.append_data(depth_vis)
|
82 |
+
else:
|
83 |
+
for i in range(frames.shape[0]):
|
84 |
+
writer.append_data(frames[i])
|
85 |
|
86 |
+
writer.close()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
video_depth_anything/video_depth.py
CHANGED
@@ -152,5 +152,5 @@ class VideoDepthAnything(nn.Module):
|
|
152 |
|
153 |
depth_list = depth_list_aligned
|
154 |
|
155 |
-
return depth_list[:org_video_len], target_fps
|
156 |
|
|
|
152 |
|
153 |
depth_list = depth_list_aligned
|
154 |
|
155 |
+
return np.stack(depth_list[:org_video_len], axis=0), target_fps
|
156 |
|