Aerial View Localization with Reinforcement Learning: Towards Emulating Search-and-Rescue

Official PyTorch implementation of the Machine Learning for Remote Sensing ICLR 2023 workshop paper Aerial View Localization with Reinforcement Learning: Towards Emulating Search-and-Rescue by Aleksis Pirinen, Anton Samuelsson, John Backsund and Kalle Åström.

arXiv | ML4RS workshop paper | SAIS 2023 version | Video | Poster

Follow-up work

Please also have a look at a follow-up to this work, titled GOMAA-Geo: GOal Modality Agnostic Active Geo-localization. GOMAA-Geo is a generalization of AiRLoc, in that it allows for different kinds of goal specifications -- whereas AiRLoc assumes goals to be specified as aerial patches, GOMAA-Geo offers zero-shot generalization to completely different goal modalities (ground-level imagery, or natural langugage text). Code for GOMAA-Geo is available here.

Installation

The code is based on Python 3.

Setup Conda environment:

conda create -n airloc
conda activate airloc
pip install -r requirements.txt

Code structure overview

All configurations of various models etcetera are set in config.py. Files related to the patch embedder is found in the folder doerchnet. Various logging (training statistics, final model weights, and so on) is sent to the folder logs. The folder data contains the data (including splits) used in the paper. Example games similar to the human performance evaluation desribed in the paper can be found in the folder human-game. When running the code the first time, first unzip all data folders.

Training

To pretrain the patch embedder in a self-supervised fashion a la Doersch et al. (2015), run from the top level folder:

python -m doerchnet.train

Various statistics (including model weights) are sent to a specific subfolder in the folder doerchnet/logs. Within that subfolder, the weights will be called doersch_embedder. Note: This repo already includes Massachusetts Buildings-pretrained patch embedder weights (both with and without segmentation as input), and the weights are already pointed to in config.py (see the flag CONFIG.RL_pretrained_doerch_net). This repo also includes xBD-pretrained embedder weights (without-sem-seg-pre-michael). Hence you may choose to skip the patch embedder pretraining step and immediately proceed to training AiRLoc (see below).

To train AiRLoc, run from the top level folder:

python -m training.train_agent

Do not forget to point CONFIG.RL_pretrained_doerch_net to the pretrained patch embedder (otherwise AiRLoc will be trained with RL from scratch, which significantly reduces the success rate of the final agent). Various statistics, including model weights, are sent to the folder logs (if this folder does not exist it will be automatically created).

To plot various statistics during training, run from the top level folder:

python -m plot_results --log-dir <folder log path> 

These statistics are stored in the folder logs.

Note: This repo already includes Massachusetts Buildings-trained AiRLoc models (both with and without segmentation as input). This repo also includes an xBD-trained AiRLoc model (2022-09-13_14-07-42-505667_pre_disaster). Hence you may choose to skip the RL-training step and immediately proceed to evaluating AiRLoc (see below).

Evaluation

To evaluate an agent on the validation set of Masschusetts Buildings, first ensure that CONFIG.MISC_dataset points to 'masa_filt'. Then run from the top level folder:

python -m eval.eval_agent --log_dir <folder log path> --seed 0 --split val 

To evaluate on the test set, change to --split test above. To evaluate on the Dubai dataset, first change CONFIG.MISC_dataset so that it points to 'dubai'. Then run from the top level folder:

python -m eval.eval_agent --log_dir <folder log path> --seed 0

Note that --log_dir refers to a folder with a trained model within the logs folder (every training run yields a subfolder in logs). It assumes model weights to be in such a subfolder, called 'final_model'.

To evaluate Priv local, do as follows:

1. Create a folder in logs, called for example PrivLocal (let's call it like that in the following).
2. Copy the doerch_embedder weights you wish to use from the desired subfolder in doerchnet/logs to the created folder (PrivLocal). Rename doerch_embedder as final_model.
3. Copy a config.py file into the PrivLocal folder, and set CONFIG.RL_agent_network = 'Agent', CONFIG.MISC_priv = True, and CONFIG.RL_LSTM_pos_emb = False.
4. Repeat the steps in the same way as described under Evaluation above. Note that --log_dir should point to PrivLocal.

To evaluate Priv local (sem seg), make sure to copy doersch_embedder weights that correspond to using a semantic segmentation channel in the input. Also set CONFIG.RL_priv_use_seg = True.

To evaluate Priv random, do as follows: To evaluate Priv local, do as follows:

1. Create a folder in logs, called for example PrivRandom (let's call it like that in the following).
2. Copy a config.py file into the PrivRandom folder, and set CONFIG.RL_agent_network = 'RandomAgent' and CONFIG.MISC_priv = True.
3. Repeat the steps in the same way as described under Evaluation above. Note that --log_dir should point to PrivRandom.

Human-controlled setup for the aerial view goal localization task

In the folder human-game there are ten zip-files, each representing a game setup similar to those provided to the human subjects that participated in the human performance evaluation mentioned in the paper. To play such a game with a graphical interface, simply unzip a file, move into the associated folder, and type:

python play.py

The above command will launch a warm-up phase for the game. To instead run a proper evaluation, run:

python play.py --real

To get information about various options, run:

python play.py -h

Create dataset

To create dataset usable by this training and evaluation setup one needs a set of satellite or drone images. Put these images in a folder called 'image' and put this folder in a surrounding folder with the intended dataset name. This folder is to be placed in the location indicated by 'CONFIG.MISC_dataset_path' in the config.py file. Then to create a split file with train/val/test split run:

python3 -m utils.dataset_utils --dataset <dataset_name> --grid-game

where <dataset_name> is the name of the folder in which the dataset is placed. Then to calculate mean and std for normalization run:

python3 -m utils.normalize_dataset --dataset <dataset_name>

where again, <dataset_name> is the name of the dataset folder. To create fixed eval splits for this dataset run:

python3 -m utils.create_split_file --dataset <dataset_name> --split <split>

where <split> is the intenteded partition of the dataset, validation is the default.

Grid sizes other than 5x5

Note that if the dataset is intended to be used with grid sizes other than the default 5x5 the config file has to be adjusted. Change the CONFIG.MISC_im_size according to the desired grid size.

Citation

If you find this implementation and/or our paper interesting or helpful, please consider citing:

@article{pirinen2022aerial,
  title={Aerial view goal localization with reinforcement learning},
  author={Pirinen, Aleksis and Samuelsson, Anton and Backsund, John and Astr{\"o}m, Kalle},
  journal={arXiv preprint arXiv:2209.03694},
  year={2022}
}