This is the reference PyTorch implementation for training and testing depth estimation models using the method described in
Real-time Monocular Depth Estimation on Embedded Systems
This code is for non-commercial use;
If you find our work useful in your research please consider citing our paper:
@article{rtmonodepth,
author = {Cheng Feng and
Zhen Chen and
Congxuan Zhang and
Weiming Hu and
Bing Li and
Liyue Ge},
title = {Real-time Monocular Depth Estimation on Embedded Systems},
booktitle = {{IEEE} International Conference on Image Processing, {ICIP} 2024,
Abu Dhabi, United Arab Emirates, October 27-30, 2024},
publisher = {{IEEE}},
year = {2024},
}
The pretrained models can be downloaded from google drive.
The code have been tested under following dependencies:
numpy=1.24.3
opencv-python=4.10.0
pillow=10.3.0
scikit-image=0.24.0
pytorch=2.3.1
torchvision=0.18.1
Following Monodepth2, we also recommend using pillow-simd
instead of pillow
for faster image preprocessing in the dataloaders.
Firstly, set the dataset root (line 10, test_simple.py) and execute python pick_test_images.py
.
Then, you can predict scaled disparity for eigen test split with:
python test_simple_full.py --image_path ./fortest/data --weight_path ./weights/RTMonoDepth/full/m_640_192/
python test_simple_s.py --image_path ./fortest/data --weight_path ./weights/RTMonoDepth/s/m_640_192/
or, if you are using a stereo-trained model, you can estimate metric depth with
python test_simple_full.py --image_path ./fortest/data --weight_path ./weights/RTMonoDepth/full/ms_640_192/ --pred_metric_depth
python test_simple_s.py --image_path ./fortest/data --weight_path ./weights/RTMonoDepth/s/ms_640_192/ --pred_metric_depth
We provide an example of comparing our small model with Fastdepth, where you can change the RTMonoDepth_s to RTMonoDesth in here to test our full model.
python compare_runtime.py
We performed our evaluation on the NVIDIA Jetson Nano and the NVIDIA Xavier NX and NVIDIA AGX Orin, using the following dependencies:
Installing PyTorch and torchvision, refer to this post: https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-11-now-available/72048
Installing torch2trt: https://github.com/NVIDIA-AI-IOT/torch2trt
You might need to increase SWAP memory for the tensorRT conversion to 4GB: https://github.com/JetsonHacksNano/resizeSwapMemory
python embedded_platform_runtime_test.py
You can also change the RTMonoDepth to RTMonoDesth_s in here to test our small model.
You can download the entire raw KITTI dataset by running:
wget -i splits/kitti_archives_to_download.txt -P kitti_data/
Then unzip with
cd kitti_data
unzip "*.zip"
cd ..
Warning: it weighs about 175GB, so make sure you have enough space to unzip too!
Our default settings expect that you have converted the png images to jpeg with this command, which also deletes the raw KITTI .png
files:
find kitti_data/ -name '*.png' | parallel 'convert -quality 92 -sampling-factor 2x2,1x1,1x1 {.}.png {.}.jpg && rm {}'
or you can skip this conversion step and train from raw png files by adding the flag --png
when training, at the expense of slower load times.
The above conversion command creates images which match our experiments, where KITTI .png
images were converted to .jpg
on Ubuntu 16.04 with default chroma subsampling 2x2,1x1,1x1
.
We found that Ubuntu 18.04 defaults to 2x2,2x2,2x2
, which gives different results, hence the explicit parameter in the conversion command.
You can also place the KITTI dataset wherever you like and point towards it with the --data_path
flag during training and evaluation.
Splits
The train/test/validation splits are defined in the splits/
folder.
By default, the code will train a depth model using Zhou's subset of the standard Eigen split of KITTI, which is designed for monocular training.
You can also train a model using the new benchmark split or the odometry split by setting the --split
flag.
Monocular training:
python train.py --model_name mono_model
Stereo training:
Our code defaults to using Zhou's subsampled Eigen training data. For stereo-only training we have to specify that we want to use the full Eigen training set – see paper for details.
python train.py --model_name RTMonoDepth_s --data_path Your_path_to_kitti_raw --num_workers=8 --batch_size=16 --scales 0 1 2 3 --num_epochs 31
Monocular + stereo training:
python train.py --model_name RTMonoDepth_ms --data_path Your_path_to_kitti_raw --num_workers=8 --batch_size=16 --scales 0 1 2 3 --num_epochs 31 --use_stereo
The code can only be run on a single GPU.
You can specify which GPU to use with the CUDA_VISIBLE_DEVICES
environment variable:
CUDA_VISIBLE_DEVICES=2 python train.py --model_name mono_model
Add the following to the training command to load an existing model for finetuning:
python train.py --model_name finetuned_mono --load_weights_folder ./log/mono_model/models/weights_19
Run python train.py -h
(or look at options.py
) to see the range of other training options, such as learning rates and ablation settings.
To prepare the ground truth depth maps run:
python export_gt_depth.py --data_path Your_path_to_kitti_raw --split eigen
python export_gt_depth.py --data_path Your_path_to_kitti_raw --split eigen_benchmark
...assuming that you have placed the KITTI dataset in the default location of ./kitti_data/
.
The following example command evaluates the pretrained models:
python evaluate_depth_full.py --data_path Your_path_to_kitti_raw --load_weights_folder ./weights/RTMonoDepth/s/m_640_192/ --eval_mono
python evaluate_depth_s.py --data_path Your_path_to_kitti_raw --load_weights_folder ./weights/RTMonoDepth/s/m_640_192/ --eval_mono
For stereo models, you must use the --eval_stereo
flag (see note below):
python evaluate_depth_full.py --data_path Your_path_to_kitti_raw --load_weights_folder ./weights/RTMonoDepth/ms/m_640_192/ --eval_stereo
python evaluate_depth_s.py --data_path Your_path_to_kitti_raw --load_weights_folder ./weights/RTMonoDepth/s/ms_640_192/ --eval_stereo
If you train your own model with our code you are likely to see slight differences to the publication results due to randomization in the weights initialization and data loading.
An additional parameter --eval_split
can be set.
The three different values possible for eval_split
are explained here:
--eval_split |
Test set size | For models trained with... | Description |
---|---|---|---|
eigen |
697 | --split eigen_zhou (default) or --split eigen_full |
The standard Eigen test files |
eigen_benchmark |
652 | --split eigen_zhou (default) or --split eigen_full |
Evaluate with the improved ground truth from the new KITTI depth benchmark |
benchmark |
500 | --split benchmark |
The new KITTI depth benchmark test files. |
Because no ground truth is available for the new KITTI depth benchmark, no scores will be reported when --eval_split benchmark
is set.
Instead, a set of .png
images will be saved to disk ready for upload to the evaluation server.
📷 Note on stereo evaluation
Our stereo models are trained with an effective baseline of 0.1
units, while the actual KITTI stereo rig has a baseline of 0.54m
. This means a scaling of 5.4
must be applied for evaluation.
In addition, for models trained with stereo supervision we disable median scaling.
Setting the --eval_stereo
flag when evaluating will automatically disable median scaling and scale predicted depths by 5.4
.
The majority of the code for this project comes from Monodepth2. We appreciate the outstanding contributions Project has made to this field.
Meanwhile, the licensing of this project is the same as that of Monodepth2.