Skip to content

Latest commit

 

History

History
70 lines (65 loc) · 3.97 KB

README.md

File metadata and controls

70 lines (65 loc) · 3.97 KB

Single-Detector

Single-Detector works on data pre-processed (output of matched filter)[DatasetA in paper], this is 4-channel data files

  • Channels 0,1,2 corresponds to R,G,B
  • Channel 3 corresponds for processed output from matched-filter as shown below

Getting Started

  • If using virtual environment, activate your virtual env

  • Setup the detector code

cd ./single_detector/src/custom-mask-rcnn-detector/mask_rcnn
python setup.py install

  • Generating training data for our detector

cd ./single_detector

Run data_generator script. This script takes 1 argument (path to "data" directory)

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

  • Training the detector

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

Training Data :  single_detector/src/custom-mask-rcnn-detector/ch4_data/train_data
Annotations   : single_detector/src/custom-mask-rcnn-detector/ch4_data/annotation_plumes.json

To train the detector, Run :

cd ./single_detector/src/custom-mask-rcnn-detector
python detector.py --mode train

The corresponsing trained weight will be stored at :

Trained weights : ./single_detector/src/custom-mask-rcnn-detector/logs

  • Testing the detector

python detector.py --mode predict --image <path to processed input file to detector>

Output is generated in the same directory

Step by Step Detection

To help with debugging and understanding the model, you can run (mask_tiling.py, json_creator.py, train_data_generator.py), that allows running the model step by step to inspect the output at each point. Here are few examples :

1. Model Architecture:

The model architecture shows data flow from left-->right, we have an input image of size ~1500 x 23000 x 4 along with a ground truth file(binary) of size 1500 x 23000 x 1. We tile each image files because of GPU memory constraint

- Ground Truth pre-processing

mask_tiling.py : creates tiles size (256 x 256)of the binary mask. Output is saved in the same datafolder. To Run it :

python mask_tiling.py ../data

json_creator.py : create a json file, considering each tile created in previous step as an individual image, then run DFS(depth first search) along the boundary of each image to create a json of boundary of plumes shapes. To run it

python json_creator.py ../data
Output location : ./single_detector/src/custom-mask-rcnn-detector/ch4_data/annotation_plumes.json

train_data_generator.py : This files creates the training data for our detector, it reads the ~1500 x 23000 x 4 (.envi format) file, process channels 0,1,2 & 3 as show in the image above --> then creates tiles of size (256 x 256 x 3) and saves them as .npy files to be used by detector. To run it

python train_data_generator.py ../data
Output Location : ./single_detector/src/custom-mask-rcnn-detector/ch4_data/train_data/