Instead of ERFNet the SegFormer implementation is used. The implementation is taken from here. It either can be trained using the pipeline provided via link or just pretrained weights can be downloaded either from NVlabs's sources or from here. The original framework is here.
Semi-Supervised Active Learning for Semantic Segmentation in Unknown Environments Using Informative Path Planning
Semantic segmentation enables robots to perceive and reason about their environments beyond geometry. Most of such systems build upon deep learning approaches. As autonomous robots are commonly deployed in initially unknown environments, pre-training on static datasets cannot always capture the variety of domains and limits the robot’s perception performance during missions. Recently, self-supervised and fully supervised active learning methods emerged to improve a robot’s vision. These approaches rely on large in-domain pre-training datasets or require substantial human labelling effort. We propose a planning method for semi-supervised active learning of semantic segmentation that substantially reduces human labelling requirements compared to fully supervised approaches. We leverage an adaptive map-based planner guided towards the frontiers of unexplored space with high model uncertainty collecting training data for human labelling. A key aspect of our approach is to combine the sparse high-quality human labels with pseudo labels automatically extracted from highly certain environment map areas. Experimental results show that our method reaches segmentation performance close to fully supervised approaches with drastically reduced human labelling effort while outperforming self-supervised approaches.
Our semi-supervised active learning approach in an unknown environment (top). We infer semantic segmentation (top-centre) and model uncertainty (top-right) and fuse both in environment maps. The robot re-plans its path (orange, top-left) to collect diverse uncertain images. After each mission, we select sparse sets of pixels for human and self-supervised labelling (bottom). Self-supervised labels are rendered from low-uncertainty semantic map regions. Human labels are queried for regions of cluttered model predictions.
During a mission, a semantic segmentation network predicts pixel-wise semantics and model uncertainties from an RGB-D image. Both are fused into an uncertainty-aware semantic environment map. Our planner guides the collection of training data for network re-training based on the robot state and map belief). After a mission, the collected data is labelled using two sources of annotations: (i) a human annotator labels a sparse set of informative pixels, and (ii) we automatically render pseudo labels from the semantic map in an uncertainty-aware fashion.
- Optional: Create a set_credentials.sh script with the following lines to enable usage of the provided Telegram and Slack experiment bots:
export TELEGRAM_CHAT_ID=<YOUR_TELEGRAM_CHAT_ID>
export TELEGRAM_TOKEN=<YOUR_TELEGRAM_TOKEN>
export SLACK_TOKEN=<YOUR_SLACK_TOKEN>
export SLACK_BOTNAME=<YOUR_SLACK_BOTNAME>
export SLACK_WEBHOOK=<YOUR_SLACK_WEBHOOK>
Learn how to set up your own Slack webhook here. Learn to setup your own Telegram bot here.
- Clone repo and initialize submodules:
git clone git@github.com:dmar-bonn/ipp-ssl.git
cd ipp-ssl
git submodule update --init
pip3 install -r bayesian_erfnet/requirements.txt
pip3 install -r requirements.txt
- Download orthomosaics, generated train-val-test splits, and pretrained Bayesian ERFNet:
./download_data.sh
- Run the semi-supervised learning pipeline:
3.1. Set environment variables in terminal:
export PYTHONPATH=$(pwd):$(pwd)/bayesian_erfnet/
source set_credentials.sh (optional)
3.2. Execute framework with orthomosaic-based simulator
python3 main.py --config_file config/<CONFIG-FILE>.yaml
Create directory for this project in current working directory and clone this repository:
mkdir tro_project
cd tro_project
git clone --recurse-submodules git@github.com:dmar-bonn/ipp-ssl.git
cd ipp-ssl/
Build docker image for this project:
cd ../../
docker build -t ssl_ipp:ral .
Download orthomosaics, generated train-val-test splits, and pretrained ERFNet:
./download_data.sh
Run planning pipeline in a docker container with NVIDIA GPU acceleration:
docker run --rm --gpus all -v $(pwd):/ipp-ssl/ -it ssl_ipp:ral bash -c "cd /ipp-ssl/ && source source_envs.sh && python3 main.py"
or without NVIDIA GPU acceleration:
docker run --rm -v $(pwd):/ipp-ssl/ -it ssl_ipp:ral bash -c "cd /ipp-ssl/ && source source_envs.sh && python3 main.py"
The pipeline executes the number of missions specified in config/config.yaml. All config files are saved to the disk. During a mission, the collected train data is saved to the disk. After each mission, the map, the planned path, and the evaluation metrics of the retrained model are saved to the disk.
If you would like to use the framework with your own pre-trained Bayesian ERFNet, orthomosaic or train-val-test data split, follow the steps below.
- Download and unpack your custom orthomosaic, e.g. Potsdam.
- Pre-train your Bayesian-ERFNet model on a semantic segmentation dataset, e.g. Cityscapes, as described here.
- Generate a train-validation-test image split of your custom orthomosaic as described in the subsection below.
- Empty the training_set/human/images and training_set/human/anno folder of your generated dataset split.
- Adapt the path_to_checkpoint in your config/<CONFIG-FILE>.yaml to your model checkpoint path.
- Adapt the path_to_orthomosaic and path_to_anno attributes in your config/<CONFIG-FILE>.yaml file to your orthomosaic RGB and Labels directories.
- Adapt the path_to_dataset in your bayesian_erfnet/agri_semantics/config/<MODEL-CONFIG-FILE>.yaml file to the dataset directory path.
- Set environment variables in terminal:
export PYTHONPATH=$(pwd):$(pwd)/bayesian_erfnet/
source set_credentials.sh (optional)
- Execute the dataset generator for orthomosaic-based simulator
python3 dataset_generator.py --config_file config/<CONFIG-FILE>.yaml --dataset_folder <DATASET-FOLDER> --num_data_samples <NUM-DATA-SAMPLES>
where is either 'training_set', 'validation_set' or 'test_set', and is the number of to be generated image-annotation data points.
In general, we follow the Python PEP 8 style guidelines. Please install black to format your python code properly. To run the black code formatter, use the following command:
black -l 120 path/to/python/module/or/package/
To optimize and clean up your imports, feel free to have a look at this solution for PyCharm.
Julius Rückin, jrueckin@uni-bonn.de, Ph.D. student at PhenoRob - University of Bonn
This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy - EXC 2070 – 390732324. Authors are with the Cluster of Excellence PhenoRob, Institute of Geodesy and Geoinformation, University of Bonn.