DNA-Depth

This project is for Deep Neighbor Layer Aggregation for Lightweight Self-Supervised Monocular Depth Estimation ICASSP 2024

This code uses PyTorch implementation for training and testing depth estimation models. The schematic of the network is shown below:

Setup

Assuming a fresh Anaconda distribution, you can install the dependencies with:

conda create -n dna python=3.6
conda activate dna
conda install six matplotlib scikit-image ipython 
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch
pip install tensorboardX==2.2 opencv-python==4.5.3.56

Our results

Train

python train.py 

Test

python evaluate_depth.py --load_weights_folder mono_640m_b0 --eval_mono --png

Trained models on KITTI

--model_name	Model resolution	KITTI abs. rel. error	delta < 1.25	FLops	params
mono_640m_b0	640 x 192	0.105	0.892	9.87	4.15
mono_640m_b1	640 x 192	0.102	0.896	10.35	6.66
mono_1024m_b0	1024 x 320	0.100	0.898	26.35	4.15
mono_1024m_b1	1024 x 320	0.097	0.902	27.58	6.66

Acknowledgement

Thanks the authors for their works:

• monodepth2
• HRNet
• DIFFNet
• EfficinetNet