Awesome
ALS4GAN
This repository contains code for the paper: Active Learning for Improved Semi Supervised Semantic Segmentation in Satellite Images which has been accepted at WACV 2022.
Abstract
Remote sensing data is crucial for applications ranging from monitoring forest fires and deforestation to tracking urbanization. Most of these tasks require dense pixel-level annotations for the model to parse visual information from limited labeled data available for these satellite images. Due to the dearth of high-quality labeled training data in this domain, recent advances have focused on semi-supervised techniques. These techniques generate pseudo-labels from a small set of labeled examples which are used to augment the labeled training set. This makes it necessary to have a highly representative and diverse labeled training set. Therefore, we propose to use an active learning-based sampling strategy to select a highly representative set of labeled training data. We demonstrate our proposed method's effectiveness on two existing semantic segmentation datasets containing satellite images: UC Merced Land Use Classification Dataset, and DeepGlobe Land Cover Classification Dataset. We report a 27% improvement in mIoU with as little as 2% labeled data using active learning sampling strategies over randomly sampling the small set of labeled training data.
Pre-requisites
Create a conda enviroment using the provided requirements.yaml file
conda env create --file=requirements.yaml
Data Peparation
UC Merced Land Use Classification Dataset
- Download the images and classification labels for the UC Merced Land Use Classification Dataset here
- Download the semantic segmentation labels for the UC Merced Land Use Classification Dataset (DLSRD Dataset) here
DeepGlobe Land Cover Classification Dataset
- Download the images and semantic segmentation labels for the DeepGlobe Land Cover Classification Dataset here
- Download the JSON file with the image classification labels for the DeepGlobe Land Cover Classification Dataset here
Training and Testing
Training the Active Learner
python tools/train_AL.py --dataset-name <DATASET_NAME>
--query-strategy <QUERY_STRATEGY>
--random-scale
--random-mirror
--data-dir <DATA_DIR>
--data-list <DATA_LIST>
--labeled-ratio <LABELED_RATIO>
--model-name <MODEL_NAME>
--num-classes <NUM_CLASSES>
--learning-rate 0.001
--num-epochs 50
--alpha <ALPHA>
--beta <BETA>
Here is an explanation of the important parameters:
dataset-name
: Can be one ofUCM
ordeepglobe
or a custom datasetquery-strategy
: Should be one ofentropy
ormargin
data-dir
: Parent directory of the datasetdata-list
: Path to the text file containing the file names of the training datasetlabeled-ratio
: Labeled ratio (value between 0 and 1, Eg: 0.05 = 5% labeled ratio)model-name
: Image classification network (Can be one ofres50
orres101
or another image classification network)num-classes
: Number of classes in the image classification networkalpha
: A value between 0 and 1 (refer the paper for details)beta
: A value between 0 and 1 (refer the paper for details)
Training the active learner would output a .npy
with the image names of the top-k most diverse samples selected by the query strategy.
Training the S4GAN Model
python tools/train_s4gan.py --dataset <DATASET_NAME>
--labeled-ratio <LABELED_RATIO>
--threshold-st <SELF_TRAINING_THRESHOLD>
--num-classes <NUM_CLASSES>
--data-dir <DATA_DIR>
--data-list <TRAIN_DATA_LIST>
--batch-size 10
--num-steps 20000
--restore-from <PRETRAINED_BACKBONE_WEIGHTS_PATH>
--save-pred-every 100
--random-seed <RANDOM_SEED>
--active-learning
--active-learning-images-array <OUTPUT OF train_al.py>
Here is an explanation of the important paramters:
dataset
: Can beucm
ordeepglobe
or a custom datasetlabeled-ratio
: Labeled ratio (value between 0 and 1, Eg: 0.05 = 5% labeled ratio)threshold-st
: Self-training theshold between 0 and 1num-classes
: Number of classes for semantic segmentationdata-dir
: Parent directory of the datasetdata-list
: Path to the text file containing the file names of the training datasetnum-steps
: Number of epochsrestore-from
: ImageNet pretrained weights that were used to initialize theresnet101
backbone of theDeepLab
generator. Download ImageNet pretrained Resnet-101(Link) and place it./pretrained_models/
random-seed
: Random seed for baseline--active-learning
: Set flag to train s4gan with the output of the active learner. Do not set this flag when training the baseline model. (Refer to the paper for more details)--active-learning-images-array
: Image array outputted by the active learner
Evaluating the S4GAN Model
python tools/auto_evaluate.py --start-eval 100
--end-eval 20000
--step-eval 100
--dataset <DATASET_NAME>
--data-dir <DATA_DIR>
--data-list <TEST_DATA_LIST>
--num-classes <NUM_CLASSES>
--checkpoint-dir <CHECKPOINT_DIR>
--input-size '320,320'
This script runs evaluation on all the weights saved during the training of s4gan and finds the epoch at which the model performs the best. Here is an explanation of the important paramters:
start-eval
: Epoch to start evaluatingend-eval
: Epoch to end evaluatingnum-classes
: Number of classes for semantic segmentationdata-dir
: Parent directory of the datasetdata-list
: Path to the text file containing the file names of the test datasetcheckpoint-dir
: Directory where the model checkpoints are stored
Acknowledgements
Parts of the code have been adapted from:
Citation
@inproceedings{desai2022active,
title={Active Learning for Improved Semi-Supervised Semantic Segmentation in Satellite Images},
author={Desai, Shasvat and Ghose, Debasmita},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={553--563},
year={2022}
}