Overview of Monocular depth estimation and BEiT

Monocular depth estimation, aiming to infer detailed depth from a single image or camera view, finds applications in fields like generative AI, 3D reconstruction, and autonomous driving. However, deriving depth from individual pixels in a single image is challenging due to the underconstrained nature of the problem. Recent advancements attribute progress to learning-based methods, particularly with MiDaS, leveraging dataset mixing and scale-and-shift-invariant loss. MiDaS has evolved with releases featuring more powerful backbones and lightweight variants for mobile use. With the rise of transformer architectures in computer vision, including those pioneered by models like ViT, there's been a shift towards using them for depth estimation. Inspired by this, MiDaS v3.1 incorporates promising transformer-based encoders alongside traditional convolutional ones, aiming for a comprehensive investigation of depth estimation techniques. The paper focuses on describing the integration of these backbones into MiDaS, providing a thorough comparison of different v3.1 models, and offering guidance on utilizing future backbones with MiDaS.

Model description

This DPT model uses the BEiT model as backbone and adds a neck + head on top for monocular depth estimation.

The previous release MiDaS v3.0 solely leverages the vanilla vision transformer ViT, MiDaS v3.1 offers additional models based on BEiT, Swin, SwinV2, Next-ViT and LeViT.

DPT 3.1 (BEiT backbone)

The highest quality depth estimation is achieved using the BEiT transformer. We provide variants such as BEiT512-L, BEiT384-L, and BEiT384-B, where the numbers signify training resolutions of 512x512 and 384x384, while the letters denote large and base models respectively. Although newer versions like BEiT v2 and BEiT-3 exist, they were not explored in our study. BEiT v2 lacked pretrained checkpoints with resolutions of 384x384 or higher, only offering them at 224x224. BEiT-3 was released after our study was completed.

DPT (Dense Prediction Transformer) model trained on 1.4 million images for monocular depth estimation. It was introduced in the paper Vision Transformers for Dense Prediction by Ranftl et al. (2021) and first released in this repository.

This model card refers specifically to BEiT512-L in the paper, and is refered to dpt-beit-large-512. A more recent paper from 2013, specifically discussing BEit, is in this paper MiDaS v3.1 – A Model Zoo for Robust Monocular Relative Depth Estimation

The model card has been written in combination by the Hugging Face team and Intel.

How to use

Be sure the to update PyTorch as Transformers as mismatches in versions can generate erros such as: "TypeError: unsupported operand type(s) for //: 'NoneType' and 'NoneType'".

As tested by this contributor, the following versions ran correctly:

import torch
import transformers
print(torch.__version__)
print(transformers.__version__)

out: '2.2.1+cpu'
out: '4.37.2'

To Install:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
 

To Use:

Here is how to use this model for zero-shot depth estimation on an image:

from transformers import DPTImageProcessor, DPTForDepthEstimation
import torch
import numpy as np
from PIL import Image
import requests

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

processor = DPTImageProcessor.from_pretrained("Intel/dpt-beit-large-512")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-beit-large-512")

# prepare image for the model
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    predicted_depth = outputs.predicted_depth

# interpolate to original size
prediction = torch.nn.functional.interpolate(
    predicted_depth.unsqueeze(1),
    size=image.size[::-1],
    mode="bicubic",
    align_corners=False,
)

# visualize the prediction
output = prediction.squeeze().cpu().numpy()
formatted = (output * 255 / np.max(output)).astype("uint8")
depth = Image.fromarray(formatted)
depth

or one can use the pipeline API:

from transformers import pipeline

pipe = pipeline(task="depth-estimation", model="Intel/dpt-beit-large-512")
result = pipe("http://images.cocodataset.org/val2017/000000181816.jpg")
result["depth"]

Quantitative Analyses

BibTeX entry and citation info

@article{DBLP:journals/corr/abs-2103-13413,
  author    = {Ren{\'{e}} Reiner Birkl, Diana Wofk, Matthias Muller},
  title     = {MiDaS v3.1 – A Model Zoo for Robust Monocular Relative Depth Estimation},
  journal   = {CoRR},
  volume    = {abs/2307.14460},
  year      = {2021},
  url       = {https://arxiv.org/abs/2307.14460},
  eprinttype = {arXiv},
  eprint    = {2307.14460},
  timestamp = {Wed, 26 Jul 2023},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2307.14460.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}