The project is about finding out the different anomalies on the surface of Mars. This project is under the category of a supervised algorithm where the data is been trained on the below set of images in order to do a small PoC.
MASK_RCNN/
├── assets/
├── datasets/ Datasets used in every model for training/
├── modules/
│ ├── anomaly/ Model 2 (Anomalies -- `crater` and `sand dune`)
│ ├── crater/ Model 1 (Anomaly -- `crater` )
│ ├── mars_anomaly/ Model 3 (Anomalies -- `slope streak, bright dune, impact ejecta, swiss cheese`)
├── mrcnn Main files for configuration, model, utilities and visualization.
├── notebooks/
│ ├── anomaly_2_classes_demo/ Demo for identification of crater and sand dune.
│ ├── crater_demo/ Demo for identification of crater.
│ ├── mars_anomaly_demo/ Demo for identification of 4 anomalies.
│ ├── 1_image_slicer.ipynb A notebook demonstrating to slice the given image.
│ ├── 2_image_annotations.ipynb Empty Notebook.
│ ├── 3_Mask_RCNN_Training.ipynb A notebook for training the model.
├── weights/
├── README.md
├── requirements.txt
├── setup.cfg
└── setup.py
- Data was annotated using VIA tool, the version was used is 1.0.6
- We used polygon shape for the outliners for the anomaly.
- For every class we have two
.jsonfile for each of the two directories consisting of 250 images for training and 50 images for validation.
- The model was trained on Google Colab with the help of GPU.
The project is done on 3 different datasets with different no of an anomaly.
-
Model 1: Identify crater which is trained on a very small dataset [1].
- The model was trained for 30 epochs having 100 batch steps per epoch is 100.
- Data was distributed into
train,validationandtest - The model was created in order to try out the object detection and segmentation for a single class of anomalies on Mars Surface.
- The accuracy isn't so good as the data to which it was trained was few.
-
Model 2: Identify crater and sand dune which is trained on a very small dataset [1].
- The model was trained for 30 epochs having 100 batch steps per epoch is 100.
- Data was distributed into
train,validationandtest. - The model was created in order to try out the object detection and segmentation for two classes (Crater and Sand-Dune) of anomalies on Mars Surface.
- The accuracy isn't so good as the data to which it was trained was few.
-
Model 3: Identify crater which is trained on a
Mars orbital image (HiRISE) labeled data set version 3dataset- The total datasets compormises of label data with the following classes,
Class-Id Anomaly Name No of Images 0 others 61k 1 crater 4.9k 2 dark dune 1.1k 3 slope streak 2.3k 4 bright dune 1.7k 5 impact ejecta 231 6 swiss cheese 1.1k 7 spider 476 - This model is trained on
3,4,5,6class-ids each of having 250 images for training (except forimpact ejecta, due to less count) and 50 for validation. - The model was trained for 30 epochs having 100 batch steps per epoch is 100.
- Data was distributed into "train", 'validation' and "test".
For a Model 1 and Model 2 we used the following two images for training, and one for testing.
- Train
- Validation
- https://photojournal.jpl.nasa.gov/catalog/PIA00179
- https://photojournal.jpl.nasa.gov/catalog/PIA00180 -Test
- https://photojournal.jpl.nasa.gov/catalog/PIA00181 For training, the image was split into the 6 small images using the library - image_slicer. For testing, we again slice the .tiff image
- Model 1 - Anomaly - crater
- Model 2 - Anomaly - crater and sand dune
- Model 3 - Anomaly - slope streak, bright dune, impact ejecta, swiss cheese
https://github.com/matterport/Mask_RCNN
[1]: