Skip to content

Latest commit

 

History

History
75 lines (63 loc) · 3.55 KB

README.md

File metadata and controls

75 lines (63 loc) · 3.55 KB
Raw

Ensemble-Detector

Ensemble detector works on raw input data (un-rectified)[Dataset B in paper]. Each datum is 432-channel/bands file. This data is first processed in sets of channels, where each set contains 50-channels/bands. Each set have an overlap of 25-channels/bands with neighbouring sets. Overview of pipeline is shown below :

Getting Started

  • If using virtual environment, activate your virtual env

  • Getting annotation data ready

cd ./ensemble_detectors

Run annotation_generator.sh. This script takes 1 argument (path to ground truth "./data/gt_data" data)

./annotation_generator.sh <path to ground truth data directory>
e.g.
(relative path) ./annotation_generator.sh '../data/gt_data'
or
(absolute path)[preferred]
./annotation_generator.sh '/home/<username>/Methane-detection-from-hyperspectral-imagery/ensemble_detectors/data/gt_data'

Annotation file saved at : Methane-detection-from-hyperspectral-imagery/ensemble_detectors/src/gt_jsonfile

  • Generating training data for the detector

This repo follows the same code execution flow as mentioned in the paper

  1. Algorithm_1_matchfilter : Algorithm 1 in paper. It runs matched filer on the raw hyperspectral data files and then runs ortho correction to rectify the data
  2. Algorithm_2_preprocess : Algorithm 2 in paper. It create training data for the detection, terrain_img_creator.py file reads the 0,1,2 bands from the 4-bands data and creates terrain image files. train_data_generator.py create sets of bands and creates tiles of images to be used in training. Each tile containes 3 bands :- 0-gray scale terrain image, 1-locally normalized output of matched filter, 2-globally normalized out of matched filter
  3. Algorithm_3_mrcnn : Algorithm 3 in paper. It contains our detector network. detector.py is our training/testing file for network
cd ./src

Run data_generator script.

./data_generator.sh

Running time of this script depends on the size of data. It does cumulative work on Algorithm_1_matchfilter & Algorithm_2_preprocess. Each hyperspectral image of dim : 22,000 x 1500 x 432 takes ~15 minutes of time to generate output

  • Training the ensemble of detector

you will find your processed training data and annotation file at below location

Training Data : ensemble_detectors/data/train_data
Annotations   : ensemble_detectors/src/gt_jsonfile
Setup detector code
cd ./ensemble_detectors/src/Algorithm_3_mrcnn/mask_rcnn
python setup.py install

To train the detector, Run

cd ./ensemble_detectors/src/Algorithm_3_mrcnn
python detector.py --mode train

The corresponsing trained weight will be stored at below location. For each detector trained, the corresponding weights are stored in the folder with name for example :

./logs/ _ _ 0-50
         |_ 25-75
         |_ 50-100
         |_ 75-125 
and so on ...

Trained weights : ./ensemble_detectors/src/Algorithm_3_mrcnn/logs

  • Testing the detector

python detector.py --mode predict --image <path to processed input file each file for each set of channels/bands>