FastDepth

This repo offers trained models and evaluation code for the FastDepth project at MIT.

Requirements
Trained Models
Evaluation
Deployment
Results
Citation

Requirements

Install PyTorch on a machine with a CUDA GPU. Our code was developed on a system running PyTorch v0.4.1.

Install the HDF5 format libraries. Files in our pre-processed datasets are in HDF5 format.

sudo apt-get update
sudo apt-get install -y libhdf5-serial-dev hdf5-tools
pip3 install h5py matplotlib imageio scikit-image opencv-python

Download the preprocessed NYU Depth V2 dataset in HDF5 format and place it under a data folder outside the repo directory. The NYU dataset requires 32G of storage space.
```
 mkdir data; cd data
 wget http://datasets.lids.mit.edu/fastdepth/data/nyudepthv2.tar.gz
 tar -xvf nyudepthv2.tar.gz && rm -f nyudepthv2.tar.gz
 cd ..
```

Trained Models

The following trained models can be found at http://datasets.lids.mit.edu/fastdepth/results/.

MobileNet-NNConv5
MobileNet-NNConv5(depthwise)
MobileNet-NNConv5(depthwise), with additive skip connections
MobileNet-NNConv5(depthwise), with additive skip connections, pruned

Our final model is mobilenet-nnconv5-skipadd-pruned, i.e. a MobileNet-NNConv5 architecture with depthwise separable layers in the decoder, with additive skip connections between the encoder and decoder, and after network pruning using NetAdapt. The other models are offered to provide insight into our approach.

When downloading, save models to a results folder outside the repo directory:

mkdir results; cd results
wget -r -np -nH --cut-dirs=2 --reject "index.html*" http://datasets.lids.mit.edu/fastdepth/results/
cd ..

Pretrained MobileNet

The model file for the pretrained MobileNet used in our model definition can be downloaded from http://datasets.lids.mit.edu/fastdepth/imagenet/.

Evaluation

This step requires a valid PyTorch installation and a saved copy of the NYU Depth v2 dataset. It is meant to be performed on a host machine with a CUDA GPU, not on an embedded platform. Deployment on an embedded device is discussed in the next section.

To evaluate a model, navigate to the repo directory and run:

python3 main.py --evaluate [path_to_trained_model]

The evaluation code will report model accuracy in terms of the delta1 metric as well as RMSE in millimeters.

Note: This evaluation code was sourced and modified from here.

Deployment

We use the TVM compiler stack to compile trained models for deployment on an NVIDIA Jetson TX2. Models are cross-compiled on a host machine and then deployed on the TX2. The tvm-compile/tuning folder in this repo contains the results of our auto-tuning the layers within our models for both the TX2 GPU and CPU. These can be used during the compilation process to achieve low model runtimes on the TX2. Outputs of TVM compilation for our trained models can be found at http://datasets.lids.mit.edu/fastdepth/results/tvm_compiled/.

On the TX2, download the trained models as explained above in the section Trained Models. The compiled model files should be located in results/tvm_compiled.

Installing the TVM Runtime

Deployment requires building the TVM runtime code on the target embedded device (that will be used solely for running a trained and compiled model). The following instructions are taken from this TVM tutorial and have been tested on a TX2 with CUDA-8.0 and LLVM-4.0 installed.

First, clone the TVM repo and modify config file:

git clone --recursive https://github.com/dmlc/tvm
cd tvm
git reset --hard ab4946c8b80da510a5a518dca066d8159473345f
git submodule update --init
cp cmake/config.cmake .

Make the following edits to the config.cmake file:

set(USE_CUDA OFF) -> set(USE_CUDA [path_to_cuda]) # e.g. /usr/local/cuda-8.0/
set(USE_LLVM OFF) -> set(USE_LLVM [path_to_llvm-config]) # e.g. /usr/lib/llvm-4.0/bin/llvm-config

Then build the runtime:

make runtime -j2

Finally, update the PYTHONPATH environment variable:

export PYTHONPATH=$PYTHONPATH:~/tvm/python

Running a Compiled Model

To run a compiled model on the device, navigate to the deploy folder and run:

python3 tx2_run_tvm.py --input-fp [path_to_input_npy_file] --output-fp [path_to_output_npy_file] --model-dir [path_to_folder_with_tvm_compiled_model_files]

Note that when running a model compiled for the GPU, a cuda argument must be specified. For instance:

python3 tx2_run_tvm.py --input-fp data/rgb.npy --output-fp data/pred.npy --model-dir ../../results/tvm_compiled/tx2_cpu_mobilenet_nnconv5dw_skipadd_pruned/
python3 tx2_run_tvm.py --input-fp data/rgb.npy --output-fp data/pred.npy --model-dir ../../results/tvm_compiled/tx2_gpu_mobilenet_nnconv5dw_skipadd_pruned/ --cuda True

Example RGB input, ground truth, and model prediction data (as numpy arrays) is provided in the data folder. To convert the .npy files into .png format, navigate into data and run python3 visualize.py.

Measuring Power Consumption

On the TX2, power consumption on the main VDD_IN rail can be measured by running the following command:

cat /sys/devices/3160000.i2c/i2c-0/0-0041/iio_device/in_power0_input

Results

Comparison against prior work. Runtimes were measured on an NVIDIA Jetson TX2 in max-N mode.

on NYU Depth v2	Input Size	MACs [G]	RMSE [m]	delta1	CPU [ms]	GPU [ms]
Eigen et al. [NIPS 2014]	228×304	2.06	0.907	0.611	300	23
Eigen et al. [ICCV 2015] (with AlexNet)	228×304	8.39	0.753	0.697	1400	96
Eigen et al. [ICCV 2015] (with VGG)	228×304	23.4	0.641	0.769	2800	195
Laina et al. [3DV 2016] (with UpConv)	228×304	22.9	0.604	0.789	2400	237
Laina et al. [3DV 2016] (with UpProj)	228×304	42.7	0.573	0.811	3900	319
Xian et al. [CVPR 2018] (with UpProj)	384×384	61.8	0.660	0.781	4400	283
This Work	224×224	0.37	0.604	0.771	37	5.6

"This Work" refers to MobileNet-NNConv5(depthwise), with additive skip connections, pruned.

Citation

If you reference our work, please consider citing the following:

@inproceedings{icra_2019_fastdepth,
	author      = {{Wofk, Diana and Ma, Fangchang and Yang, Tien-Ju and Karaman, Sertac and Sze, Vivienne}},
	title       = {{FastDepth: Fast Monocular Depth Estimation on Embedded Systems}},
	booktitle   = {{IEEE International Conference on Robotics and Automation (ICRA)}},
	year        = {{2019}}
}

dwofk/fast-depth