Fast Scene Understanding
Torch implementation for simultaneous image segmentation, instance segmentation and single image depth. Two videos can be found here and here.
Example:
If you use this code for your research, please cite our papers:
Fast Scene Understanding for Autonomous Driving
Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans and Luc Van Gool
Published at "Deep Learning for Vehicle Perception", workshop at the IEEE Symposium on Intelligent Vehicles 2017
and
Semantic Instance Segmentation with a Discriminative Loss Function
Bert De Brabandere, Davy Neven and Luc Van Gool
Published at "Deep Learning for Robotic Vision", workshop at CVPR 2017
Setup
Prerequisites
Torch dependencies:
Data dependencies:
Download Cityscapes and run the script createTrainIdLabelImgs.py
and createTrainIdInstanceImgs.py
to create annotations based on the training labels. Make sure that the folder is named cityscapes
Afterwards create following txt files:
cd CITYSCAPES_FOLDER
ls leftImg8bit/train/*/*.png > trainImages.txt
ls leftImg8bit/val/*/*.png > valImages.txt
ls gtFine/train/*/*labelTrainIds.png > trainLabels.txt
ls gtFine/val/*/*labelTrainIds.png.png > valLabels.txt
ls gtFine/train/*/*instanceTrainIds.png > trainInstances.txt
ls gtFine/val/*/*instanceTrainIds.png.png > valInstances.txt
ls disparity/train/*/*.png > trainDepth.txt
ls disparity/val/*/*.png.png > valDepth.txt
Download pretrained model
To download both the pretrained segmentation model for training and a trained model for testing, run:
sh download.sh
Test pretrained model
To test the pretrained model, make sure you downloaded both the pretrained model and the Cityscapes dataset (+ scripts and txt files, see above). After, run:
qlua test.lua -data_root CITYSCAPES_ROOT
with CITYSCAPES_ROOT the folder where cityscapes is located. For other options, see test_opts.lua
Train your own model
To train your own model, run:
qlua main.lua -data_root CITYSCAPES_ROOT -save true -directory PATH_TO_SAVE
For other options, see opts.lua