To train on your computer:
install tensorflow
sudo pip3 install utm python-opencv opencv opencv-python scipy h5py matplotlib python3-tk scipy
sudo apt install python3-tk python-scipy libpcl1.7 build-essential cmake libopencv-dev libpcl-dev libproj-dev
cd rgbd/build
cmake ..
./rgbd_dataset 8 0.60 0.05 60
cd ../..
bash download/download_all.sh
python scripts/gen.py
Check which setting number to use in src/constants.py
src/train.py SETTING_NUMBER NUMBER_OF_STEPS
Train all using
src/train.py all NUMBER_OF_STEPS
The dataset number is 0 for Lip6Indoor and 1 for Lip6Outdoor Again, check which setting number to use in src/constants.py
python scripts/test_lipdataset.py SETTING_NUMBER DATASET_NUMBER
To test all models, run the command:
python scripts/test_lipdataset.py all DATASET_NUMBER
Process SceneNN
Add way to download Streetview
Train all networks and test them
Staging to speed up training
Fix image stat calculations
Use ESC to exit the viewer
Varying the ratio may make it train more quickly. Investigate more. it train more quickly. Investigate more.