KittiSeg performs segmentation of roads by utilizing an FCN based model. The model achieved first place on the Kitti Road Detection Benchmark at submission time. Check out our paper for a detailed model description.
The model is designed to perform well on small datasets. The training is done using just 250 densely labelled images. Despite this a state-of-the art MaxF1 score of over 96% is achieved. The model is usable for real-time application. Inference can be performed at the impressive speed of 95ms per image.
The repository contains code for training, evaluating and visualizing semantic segmentation in TensorFlow. It is build to be compatible with the TensorVision back end which allows to organize experiments in a very clean way. Also check out KittiBox a similar projects to perform state-of-the art detection. And finally the MultiNet repository contains code to jointly train segmentation, classification and detection. KittiSeg and KittiBox are utilized as submodules in MultiNet.
The code requires Tensorflow 1.0, python 2.7 as well as the following python libraries:
- matplotlib
- numpy
- Pillow
- scipy
- commentjson
Those modules can be installed using: pip install numpy scipy pillow matplotlib commentjson
or pip install -r requirements.txt
.
- Clone this repository:
git clone https://github.com/MarvinTeichmann/KittiSeg.git
- Initialize all submodules:
git submodule update --init --recursive
- [Optional] Download Kitti Road Data:
- Retrieve kitti data url here: http://www.cvlibs.net/download.php?file=data_road.zip
- Call
python download_data.py --kitti_url URL_YOU_RETRIEVED
Running the model using demo.py
does not require you to download kitti data (step 3). Step 3 is only required if you want to train your own model using train.py
or bench a model agains the official evaluation score evaluate.py
. Also note, that I recommend using download_data.py
instead of downloading the data yourself. The script will also extract and prepare the data. See Section Manage data storage if you like to control where the data is stored.
- Pull all patches:
git pull
- Update all submodules:
git submodule update --init --recursive
If you forget the second step you might end up with an inconstant repository state. You will already have the new code for KittiSeg but run it old submodule versions code. This can work, but I do not run any tests to verify this.
Run: python demo.py --input_image data/demo/demo.png
to obtain a prediction using demo.png as input.
Run: python evaluate.py
to evaluate a trained model.
Run: python train.py --hypes hypes/KittiSeg.json
to train a model using Kitti Data.
If you like to understand the code, I would recommend looking at demo.py first. I have documented each step as thoroughly as possible in this file.
KittiSeg allows to separate data storage from code. This is very useful in many server environments. By default, the data is stored in the folder KittiSeg/DATA
and the output of runs in KittiSeg/RUNS
. This behaviour can be changed by setting the bash environment variables: $TV_DIR_DATA
and $TV_DIR_RUNS
.
Include export TV_DIR_DATA="/MY/LARGE/HDD/DATA"
in your .profile
and the all data will be downloaded to /MY/LARGE/HDD/DATA/data_road
. Include export TV_DIR_RUNS="/MY/LARGE/HDD/RUNS"
in your .profile
and all runs will be saved to /MY/LARGE/HDD/RUNS/KittiSeg
KittiSeg helps you to organize large number of experiments. To do so the output of each run is stored in its own rundir. Each rundir contains:
output.log
a copy of the training output which was printed to your screentensorflow events
tensorboard can be run in rundirtensorflow checkpoints
the trained model can be loaded from rundir[dir] images
a folder containing example output images.image_iter
controls how often the whole validation set is dumped[dir] model_files
A copy of all source code need to build the model. This can be very useful of you have many versions of the model.
To keep track of all the experiments, you can give each rundir a unique name with the --name
flag. The --project
flag will store the run in a separate subfolder allowing to run different series of experiments. As an example, python train.py --project batch_size_bench --name size_5
will use the following dir as rundir: $TV_DIR_RUNS/KittiSeg/batch_size_bench/size_5_KittiSeg_2017_02_08_13.12
.
The flag --nosave
is very useful to not spam your rundir.
The model is controlled by the file hypes/KittiSeg.json
. Modifying this file should be enough to train the model on your own data and adjust the architecture according to your needs. A description of the expected input format can be found here.
For advanced modifications, the code is controlled by 5 different modules, which are specified in hypes/KittiSeg.json
.
"model": {
"input_file": "../inputs/kitti_seg_input.py",
"architecture_file" : "../encoder/fcn8_vgg.py",
"objective_file" : "../decoder/kitti_multiloss.py",
"optimizer_file" : "../optimizer/generic_optimizer.py",
"evaluator_file" : "../evals/kitti_eval.py"
},
Those modules operate independently. This allows easy experiments with different datasets (input_file
), encoder networks (architecture_file
), etc. Also see TensorVision for a specification of each of those files.
KittiSeg is build on top of the TensorVision TensorVision backend. TensorVision modularizes computer vision training and helps organizing experiments.
To utilize the entire TensorVision functionality install it using
$ cd KittiSeg/submodules/TensorVision
$ python setup.py install
Now you can use the TensorVision command line tools, which includes:
tv-train --hypes hypes/KittiSeg.json
trains a json model.
tv-continue --logdir PATH/TO/RUNDIR
trains the model in RUNDIR, starting from the last saved checkpoint. Can be used for fine tuning by increasing max_steps
in model_files/hypes.json
.
tv-analyze --logdir PATH/TO/RUNDIR
evaluates the model in RUNDIR
Here are some Flags which will be useful when working with KittiSeg and TensorVision. All flags are available across all scripts.
--hypes
: specify which hype-file to use
--logdir
: specify which logdir to use
--gpus
: specify on which GPUs to run the code
--name
: assign a name to the run
--project
: assign a project to the run
--nosave
: debug run, logdir will be set to debug
In addition the following TensorVision environment Variables will be useful:
$TV_DIR_DATA
: specify meta directory for data
$TV_DIR_RUNS
: specify meta directory for output
$TV_USE_GPUS
: specify default GPU behaviour.
On a cluster it is useful to set $TV_USE_GPUS=force
. This will make the flag --gpus
mandatory and ensure, that run will be executed on the right GPU.
Please have a look into the FAQ. Also feel free to open an issue to discuss any questions not covered so far.
If you benefit from this code, please cite our paper:
@article{teichmann2016multinet,
title={MultiNet: Real-time Joint Semantic Reasoning for Autonomous Driving},
author={Teichmann, Marvin and Weber, Michael and Zoellner, Marius and Cipolla, Roberto and Urtasun, Raquel},
journal={arXiv preprint arXiv:1612.07695},
year={2016}
}