1、MiDaS

参考：
https://github.com/isl-org/MiDaS
https://pytorch.org/hub/intelisl_midas_v2/
https://colab.research.google.com/github/pytorch/pytorch.github.io/blob/master/assets/hub/intelisl_midas_v2.ipynb#scrollTo=5A32CL3tocrZ

代码

import cv2
import torch
import urllib.request

import matplotlib.pyplot as plt

url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
urllib.request.urlretrieve(url, filename)


model_type = "DPT_Large"     # MiDaS v3 - Large     (highest accuracy, slowest inference speed)
#model_type = "DPT_Hybrid"   # MiDaS v3 - Hybrid    (medium accuracy, medium inference speed)
#model_type = "MiDaS_small"  # MiDaS v2.1 - Small   (lowest accuracy, highest inference speed)

midas = torch.hub.load("intel-isl/MiDaS", model_type)

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
midas.to(device)
midas.eval()

midas_transforms = torch.hub.load("intel-isl/MiDaS", "transforms")

if model_type == "DPT_Large" or model_type == "DPT_Hybrid":
    transform = midas_transforms.dpt_transform
else:
    transform = midas_transforms.small_transform

img = cv2.imread(filename)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

input_batch = transform(img).to(device)

with torch.no_grad():
    prediction = midas(input_batch)

    prediction = torch.nn.functional.interpolate(
        prediction.unsqueeze(1),
        size=img.shape[:2],
        mode="bicubic",
        align_corners=False,
    ).squeeze()

output = prediction.cpu().numpy()

plt.imshow(output)
# plt.show()

视频

import cv2
import torch
import numpy as np
import torch.nn.functional as F

# 假设你已经有一个训练好的模型 'midas'
# midas = torch.load('path_to_your_model')
# midas.eval()

# 加载视频
video_path = '20240807_024802.mp4'
cap = cv2.VideoCapture(video_path)

# 获取视频的宽度和高度
frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

# 获取视频的帧率
fps = cap.get(cv2.CAP_PROP_FPS)

# 创建 VideoWriter 对象
output_path = 'output_video2.mp4'
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(output_path, fourcc, fps, (frame_width, frame_height), isColor=True)

while cap.isOpened():
    ret, frame = cap.read()
    if not ret:
        break

    # 将帧转换为模型输入格式
    img = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)

    input_batch = transform(img).to(device)

    with torch.no_grad():
        prediction = midas(input_batch)

        prediction = torch.nn.functional.interpolate(
            prediction.unsqueeze(1),
            size=frame.shape[:2],
            mode="bicubic",
            align_corners=False,
        ).squeeze()

    output = prediction.cpu().numpy()

    # 可视化结果
    output = (output - output.min()) / (output.max() - output.min())  # 归一化到 [0, 1]
    # output = (output * 255).astype(np.uint8)
    # output = cv2.cvtColor(output, cv2.COLOR_GRAY2BGR)

    output = cv2.applyColorMap((output * 255).astype(np.uint8), cv2.COLORMAP_JET)

    # 写入视频
    out.write(output)



cap.release()
out.release()
cv2.destroyAllWindows()

2、ZoeDepth

https://huggingface.co/spaces/shariqfarooq/ZoeDepth

https://github.com/isl-org/ZoeDepth
https://colab.research.google.com/github/isl-org/ZoeDepth/blob/main/notebooks/ZoeDepth_quickstart.ipynb#scrollTo=OJ9bY7rrVuAq

报错：ZoeDepth: Unexpected key(s) in state_dict: “core.core.pretrained.model.blocks.0.attn.relative_position_index”
解决： pip install timm==0.6.7

代码：
git clone https://github.com/isl-org/ZoeDepth.git
cd ZoeDepth

import torch
from zoedepth.utils.misc import get_image_from_url, colorize
from PIL import Image
import matplotlib.pyplot as plt


zoe = torch.hub.load(".", "ZoeD_N", source="local", pretrained=True, version='ZoeD_N-Nov13-2023')
zoe = zoe.to('cuda')

#@title Predicting depth from a url image
img_url = "http://static1.squarespace.com/static/6213c340453c3f502425776e/62f2452bc121595f4d87c713/62f3c63c5eec2b12a333f851/1661442296756/Screenshot+2022-08-10+at+15.55.27.png?format=1500w" #@param {type:"string"}
img = get_image_from_url(img_url)
depth = zoe.infer_pil(img)


colored_depth = colorize(depth)
fig, axs = plt.subplots(1,2, figsize=(15,7))
for ax, im, title in zip(axs, [img, colored_depth], ['Input', 'Predicted Depth']):
  ax.imshow(im)
  ax.axis('off')
  ax.set_title(title)

本地图片

img = Image.open("/content/loong.png").convert("RGB")  # load
# depth_numpy = zoe.infer_pil(image)  # as numpy

depth = zoe.infer_pil(img)


colored_depth = colorize(depth)
fig, axs = plt.subplots(1,2, figsize=(15,7))
for ax, im, title in zip(axs, [img, colored_depth], ['Input', 'Predicted Depth']):
  ax.imshow(im)
  ax.axis('off')
  ax.set_title(title)

看着原图与深度图尺寸一致

本地运行完整代码：

import torch

from zoedepth.utils.misc import get_image_from_url, colorize
from PIL import Image
import matplotlib.pyplot as plt

# Zoe_N
model_zoe_n = torch.hub.load(r".", "ZoeD_N", source="local", pretrained=True)

img = Image.open(r"C:\Users\loong\Downloads\loong.png").convert("RGB")  # load
# depth_numpy = zoe.infer_pil(image)  # as numpy

depth = model_zoe_n.infer_pil(img)

print(depth.shape,depth)

colored_depth = colorize(depth)
fig, axs = plt.subplots(1,2, figsize=(15,7))
for ax, im, title in zip(axs, [img, colored_depth], ['Input', 'Predicted Depth']):
  ax.imshow(im)
  ax.axis('off')
  ax.set_title(title)

# Save the figure
output_filename = "output_image.png"
plt.savefig(output_filename)

# Show the figure
plt.show()

彩色

import numpy as np

img = Image.open("/content/loong.png").convert("RGB")  # load

depth = zoe.infer_pil(img)


# 归一化深度图
depth_normalized = (depth - depth.min()) / (depth.max() - depth.min())
# 反转颜色映射
depth_normalized = 1 - depth_normalized
# 应用颜色映射
colored_depth = plt.get_cmap('jet')(depth_normalized)
# 将 RGBA 转换为 RGB
colored_depth = (colored_depth[:, :, :3] * 255).astype(np.uint8)
    
fig, axs = plt.subplots(1, 2, figsize=(15, 7))
for ax, im, title in zip(axs, [img, colored_depth], ['Input', 'Predicted Depth']):
    ax.imshow(im)
    ax.axis('off')
    ax.set_title(title)

plt.show()

transformers加载使用
https://huggingface.co/docs/transformers/main/en/model_doc/zoedepth

from transformers import pipeline
from PIL import Image
import requests

# url = "http://images.cocodataset.org/val2017/000000039769.jpg"
# image = Image.open(requests.get(url, stream=True).raw)


image = Image.open(r"C:\Users\loong\Downloads\right1.png")

pipe = pipeline(task="depth-estimation", model="Intel/zoedepth-nyu-kitti")
result = pipe(image)
depth = result["depth"]

3、Depth-Anything

https://github.com/DepthAnything/Depth-Anything-V2

from transformers import pipeline
from PIL import Image

pipe = pipeline(task="depth-estimation", model="depth-anything/Depth-Anything-V2-Small-hf")
image = Image.open(r"C:\Users\loong\Downloads\right1.png")
result = pipe(image)
depth = result["depth"]