dpt

Adabins

Vision Transformer for Dense Prediction

Introduction

Official Repo

Abstract

We introduce dense vision transformers, an architecture that leverages vision transformers in place of convolutional networks as a backbone for dense prediction tasks. We assemble tokens from various stages of the vision transformer into image-like representations at various resolutions and progressively combine them into full-resolution predictions using a convolutional decoder. The transformer backbone processes representations at a constant and relatively high resolution and has a global receptive field at every stage. These properties allow the dense vision transformer to provide finer-grained and more globally coherent predictions when compared to fully-convolutional networks. Our experiments show that this architecture yields substantial improvements on dense prediction tasks, especially when a large amount of training data is available. For monocular depth estimation, we observe an improvement of up to 28% in relative performance when compared to a state-of-the-art fully-convolutional network. When applied to semantic segmentation, dense vision transformers set a new state of the art on ADE20K with 49.02% mIoU. We further show that the architecture can be fine-tuned on smaller datasets such as NYUv2, KITTI, and Pascal Context where it also sets the new state of the art. Our models are available at this https URL.

Framework

Citation

@inproceedings{ranftl2021vision,
  title={Vision transformers for dense prediction},
  author={Ranftl, Ren{\'e} and Bochkovskiy, Alexey and Koltun, Vladlen},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={12179--12188},
  year={2021}
}

Usage

To use other repositories' pre-trained models, it is necessary to convert keys.

We provide a script vit2depth.py in the tools directory to convert the key of models from timm to MMSegmentation style.

python tools/model_converters/vit2depth.py ${PRETRAIN_PATH} ${STORE_PATH}

E.g.

python tools/model_converters/vit2depth.py https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth pretrain/jx_vit_base_p16_224-80ecf9dd.pth

This script convert model from PRETRAIN_PATH and store the converted model in STORE_PATH.

Results and models

This is a simple implementation. Only model structure can be aligned with original paper. More experiments about training settings or loss functions are needed to be done.

We have achieved better results compared with results presented in our paper DepthFormer by conducting more carefully designed tricks.

In our reproduction, we utilize the standard ImageNet pre-trained ViT-Base instead of the ADE20K pre-trained model in the original paper, which is fairer to compare with other monodepth methods. We find it seems that with direct training on a small dataset (like KITTI and NYU), the model tends to be overfitting and cannot achieve satisfying results.

KITTI

Method	Backbone	Train Epoch	Abs Rel (+flip)	RMSE (+flip)	Config	Download	GPUs
DPT	ViT-Base	24	0.073	2.604	config	log | model	8 V100s

NYU

Method	Backbone	Train Epoch	Abs Rel (+flip)	RMSE (+flip)	Config	Download	GPUs
DPT	ViT-Base	24	0.135	0.413	config	log | model	8 V100s