DepthAnythingV2

Metric Depth Estimation Benchmarking and Evaluation

This subfolder contains scripts for training and evaluation for the task of Metric Depth Estimation for WildCross, using DepthAnythingV2.

Setup

Dataset

To download the WildCross Dataset, follow the instructions in the root directory of this repository. By default this repository will use the full sized images (images) and point clouds (depth) for training and evaluation.

Environment

We provide an environment file to set up the necessary python environment for training and evaluation using mamba. The environment can be installed by running the following command out of this directory:

mamba env create -f environment.yaml

Checkpoints

We provide download links for both the KITTI pre-trained and WildCross fine-tuned checkpoints for each network backbone:

Backbone	Checkpoint	Link
ViT-s	Pre-trained	Download
	Fine-tuned	Download
ViT-b	Pre-trained	Download
	Fine-tuned	Download
ViT-l	Pre-trained	Download
	Fine-tuned	Download

Training

To train on the WildCross dataset, firstly edit the value of the wildcross_root variable on line 74 of train.py to the directory containing the WildCross dataset on your machine. Then, you can train the network by running train.py as follows:

export PYTHONPATH=$PWD:$PYTHONPATH 
GPUS=$NUM_GPUS

python3 -m torch.distributed.launch \
    --nproc_per_node=$GPUS \
    --nnodes 1 \
    --node_rank=0 \
    --master_addr=localhost \
    --master_port=20596 \
    train.py \
        --encoder $ENCODER \
        --save-path /path/to/save/directory \
        --pretrained-from /path/to/pretrained/checkpoint \

Where:

• GPUS refers to the number of GPUS to be used for training
• --encoder refers to the backbone encoder to be used. We recommend using vits, vitb or vitl for small, base or large Visual Transformers respectively.
• --save-path refers to the directory to save training logs and checkpoints in
• --pretrained-from refers to the checkpoint to use as the initialisation for training. We strongly recommend using the KITTI pre-trained model to intialise the network before fine-tuning on WildCross.

Evaluation

To evaluate a given checkpoint, run eval.py as follows:

export PYTHONPATH=$PWD:$PYTHONPATH
python3 eval.py \
    --encoder $ENCODER \
    --checkpoint /path/to/checkpoint.pth \

Where --encoder is the same as used in the training script, and here --checkpoint refers to the checkpoint being evaluated.

Acknowledgements

We would like again to acknowledge the authors of DepthAnythingV2 and the maintainers of the open source implementation found at https://github.com/DepthAnything/Depth-Anything-V2, which has formed the basis for the code in this repository.