Fast CNN Stereo Depth Estimation through Embedded GPU Device

This is the implementation of our paper"Fast CNN Stereo Depth Estimation through Embedded GPU Device"

---

Results

Model	Dataset	EPE	Err > 3
Default	Kitti2012	1.80	0.11

---

Reproduce results (Training in Desktop)

Docker (training on x64 arch)

./docker/launch.sh

# Sceneflow training
python cli.py festereo-train --num_workers 16 --max_epochs 20 --min_epochs 1 --patience 100 --lr 5e-3 --save_top_k 20

# Kitti2012 (using sceneflow pretrained)
python cli.py festereo-train --num_workers 16 --max_epochs 300 --min_epochs 200 --patience 100 --lr 5e-3 --dataset kitti2012 --pretrained [path]/sceneflow_ckpt_epoch_19.ckpt --scheduler plateau 

Pretrained networks

• Sceneflow checkpoint epoch 19
• Kitti2012

---

Reproduce results (Inference in Jetson Jetpack 4.3)

Install Pytorch 1.3

Follow instructions from link

Install Python requirements

pip3 install -r requirements.txt

Install tensorrt

sudo apt-get install libprotobuf* protobuf-compiler ninja-build
git clone https://github.com/ngunsu/torch2trt.git
cd torch2trt
sudo python3 setup.py install --plugins

Install cuda kernel

cd cuda && python3 setup.py install --user

Benchmark speed

sudo nvpmodel -m 0 && sudo jetson_clocks
python3 model/default.py --benchmark --tensorrt

For more options python3 model/default.py --help

---

Citation

@article{Aguilera_2020,
	doi = {10.3390/s20113249},
	url = {https://doi.org/10.3390%2Fs20113249},
	year = 2020,
	month = {jun},
	publisher = {{MDPI} {AG}},
	volume = {20},
	number = {11},
	pages = {3249},
	author = {Cristhian A. Aguilera and Cristhian Aguilera and Crist{\'{o}}bal A. Navarro and Angel D. Sappa},
	title = {Fast {CNN} Stereo Depth Estimation through Embedded {GPU} Devices},
	journal = {Sensors}
}