Feature Hallucination Using Conditional GAN

Implementation of the paper - An Adversarial Approach to Discriminative Modality Hallucination for Remote Sensing Data - IEEE ICCV - CROMOL 2019 in TensorFlow

Uses a conditional GAN with novel discriminator of 2C classes (corresponsing to C real and fake classes each) to hallucinate missing features for hyperspectral or multimodal remote sensing data using available sensor derived features.

Generator takes latent dimensions as well as conditional input from available modality to generate missing features. GAN was trained using tips from Soumith Chintala's GANHacks.

Deep CNNs have been utilized as feature extractors for application of GAN, and further classification. Hallucinated features are concatenated with original features and sent to DNN classifier.

Ablation studies done with Indian Pines dataset.

Poster and Architecture

t-SNE Plots

Without sensor abnormality	With missing bands due to failures

Datasets

• Indian Pines Corrected
  • http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes
• Pavia University Scene
  • http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes
• Houston Fusion
  • http://hyperspectral.ee.uh.edu/?page_id=1075
• Multimodal Dataset
  • Uses Multispectral and complementary Panchromatic data.
  • Dataset is not publicly available.

Results

Train Test Split Experiments

(Temperature 2, 50-50 Band Split)

Split	75-25	50-50	25-75	10-90
Teacher Network	97.54194303550526%	90.71219512195122%	70.28749837697956%	69.51761517809012%
Student Network	93.79633242294186%	92.13658536701668%	80.57759854532085%	71.92411924248465%

Temperature Experiments

(25-75 Train Test Split, 50-50 Band Split)

Temperature	1	2	3	5	10
Teacher Network	69.4289059482037%	85.625081307652%	72.19981788053645%	81.63132561622491%	57.47365682630959%
Student Network	77.31234552461086%	95.4728762846364%	79.84909587848074%	84.75347990733494%	64.39443216129073%

Band Split Experiments (Hallucinated-Real)

(Temperature 2, 25-75 Train Test Split)

Split	25-75	50-50	75-25	90-10
Teacher Network	55.60036425191497%	85.625081307652%	62.29998699334999%	64.70664758799801%
Student Network	56.70612723088876%	95.4728762846364%	86.09340444906985%	81.43619097332131%

Usage

Run the files in following order, modifying the directory of dataset -

python twostreamdistillation.py
python train.py

The file testmodels.py is used to evaluate the optimal model for generator.

python testmodels.py
python test.py

Citation

If this paper and/or this code is useful for your research, please consider citing us -

@inproceedings{pande2019adversarial,
  title={An Adversarial Approach to Discriminative Modality Distillation for Remote Sensing Image Classification},
  author={Pande, Shivam and Banerjee, Avinandan and Kumar, Saurabh and Banerjee, Biplab and Chaudhuri, Subhasis},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision Workshops},
  pages={0--0},
  year={2019}
}