File size: 7,448 Bytes
5555f5e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 | # ruff: noqa
# fmt: off
"""Video Depth Anything, vendored verbatim from DepthAnything/Video-Depth-Anything @ 4f5ae23172ba60fd7bc11ef671cca678842c7072: the files below
concatenated, package-internal imports replaced by relative imports of the other vendored files (generated by scripts/model_ports/common/vendor_flat.py;
do not edit by hand). Code license: Apache-2.0."""
from .vda_dinov2 import DINOv2
from .vda_dpt import DPTHeadTemporal, NormalizeImage, PrepareForNet, Resize, compute_scale_and_shift, get_interpolate_frames
# ==============================================================================================================
# Video-Depth-Anything/video_depth_anything/video_depth_stream.py
# ==============================================================================================================
# Copyright (2025) Bytedance Ltd. and/or its affiliates
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
import torch.nn.functional as F
import torch.nn as nn
from torchvision.transforms import Compose
import cv2
import numpy as np
# infer settings, do not change
INFER_LEN = 32
OVERLAP = 10
INTERP_LEN = 8
class VideoDepthAnything(nn.Module):
def __init__(
self,
encoder='vitl',
features=256,
out_channels=[256, 512, 1024, 1024],
use_bn=False,
use_clstoken=False,
num_frames=32,
pe='ape'
):
super(VideoDepthAnything, self).__init__()
self.intermediate_layer_idx = {
'vits': [2, 5, 8, 11],
"vitb": [2, 5, 8, 11],
'vitl': [4, 11, 17, 23]
}
self.encoder = encoder
self.pretrained = DINOv2(model_name=encoder)
self.head = DPTHeadTemporal(self.pretrained.embed_dim, features, use_bn, out_channels=out_channels, use_clstoken=use_clstoken, num_frames=num_frames, pe=pe)
self.transform = None
self.frame_id_list = []
self.frame_cache_list = []
self.gap = (INFER_LEN - OVERLAP) * 2 - 1 - (OVERLAP - INTERP_LEN)
assert self.gap == 41
self.id = -1
def forward(self, x):
return self.forward_depth(self.forward_features(x), x.shape)[0]
def forward_features(self, x):
features = self.pretrained.get_intermediate_layers(x.flatten(0,1), self.intermediate_layer_idx[self.encoder], return_class_token=True)
return features
def forward_depth(self, features, x_shape, cached_hidden_state_list=None):
B, T, C, H, W = x_shape
patch_h, patch_w = H // 14, W // 14
depth, cur_cached_hidden_state_list = self.head(features, patch_h, patch_w, T, cached_hidden_state_list=cached_hidden_state_list)
depth = F.interpolate(depth, size=(H, W), mode="bilinear", align_corners=True)
depth = F.relu(depth)
return depth.squeeze(1).unflatten(0, (B, T)), cur_cached_hidden_state_list # return shape [B, T, H, W]
def infer_video_depth_one(self, frame, input_size=518, device='cuda', fp32=False):
self.id += 1
if self.transform is None: # first frame
# Initialize the transform
frame_height, frame_width = frame.shape[:2]
self.frame_height = frame_height
self.frame_width = frame_width
ratio = max(frame_height, frame_width) / min(frame_height, frame_width)
if ratio > 1.78: # we recommend to process video with ratio smaller than 16:9 due to memory limitation
input_size = int(input_size * 1.777 / ratio)
input_size = round(input_size / 14) * 14
self.transform = Compose([
Resize(
width=input_size,
height=input_size,
resize_target=False,
keep_aspect_ratio=True,
ensure_multiple_of=14,
resize_method='lower_bound',
image_interpolation_method=cv2.INTER_CUBIC,
),
NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
PrepareForNet(),
])
# Inference the first frame
cur_list = [torch.from_numpy(self.transform({'image': frame.astype(np.float32) / 255.0})['image']).unsqueeze(0).unsqueeze(0)]
cur_input = torch.cat(cur_list, dim=1).to(device)
with torch.no_grad():
with torch.autocast(device_type=device, enabled=(not fp32)):
cur_feature = self.forward_features(cur_input)
x_shape = cur_input.shape
depth, cached_hidden_state_list = self.forward_depth(cur_feature, x_shape)
depth = depth.to(cur_input.dtype)
depth = F.interpolate(depth.flatten(0,1).unsqueeze(1), size=(frame_height, frame_width), mode='bilinear', align_corners=True)
# Copy multiple cache to simulate the windows
self.frame_cache_list = [cached_hidden_state_list] * INFER_LEN
self.frame_id_list.extend([0] * (INFER_LEN - 1))
new_depth = depth[0][0].cpu().numpy()
else:
frame_height, frame_width = frame.shape[:2]
assert frame_height == self.frame_height
assert frame_width == self.frame_width
# infer feature
cur_input = torch.from_numpy(self.transform({'image': frame.astype(np.float32) / 255.0})['image']).unsqueeze(0).unsqueeze(0).to(device)
with torch.no_grad():
with torch.autocast(device_type=device, enabled=(not fp32)):
cur_feature = self.forward_features(cur_input)
x_shape = cur_input.shape
cur_list = self.frame_cache_list[0:2] + self.frame_cache_list[-INFER_LEN+3:]
'''
cur_id = self.frame_id_list[0:2] + self.frame_id_list[-INFER_LEN+3:]
print(f"cur_id: {cur_id}")
'''
assert len(cur_list) == INFER_LEN - 1
cur_cache = [torch.cat([h[i] for h in cur_list], dim=1) for i in range(len(cur_list[0]))]
# infer depth
with torch.no_grad():
with torch.autocast(device_type=device, enabled=(not fp32)):
depth, new_cache = self.forward_depth(cur_feature, x_shape, cached_hidden_state_list=cur_cache)
depth = depth.to(cur_input.dtype)
depth = F.interpolate(depth.flatten(0,1).unsqueeze(1), size=(frame_height, frame_width), mode='bilinear', align_corners=True)
depth_list = [depth[i][0].cpu().numpy() for i in range(depth.shape[0])]
new_depth = depth_list[-1]
self.frame_cache_list.append(new_cache)
# adjust the sliding window
self.frame_id_list.append(self.id)
if self.id + INFER_LEN > self.gap + 1:
del self.frame_id_list[1]
del self.frame_cache_list[1]
return new_depth
|