Skip to content

Latest commit

 

History

History
55 lines (43 loc) · 3.19 KB

README.md

File metadata and controls

55 lines (43 loc) · 3.19 KB
Raw

Gulf of Mexico fish lidar classification

This directory contains code used to process the Gulf of Mexico dataset. The code can be split into several groups:

  • data preprocessing
  • creating training/test sets
  • parameter tuning
  • classifier training
  • classifier testing and results

Data preprocessing

script_extract_data_from_day.m converts the original PNG files into mat files. create_processed_mat_files.m preprocesses the data.

Creating training/test sets

create_first_day_training_data.m does a train/test split by using the first day of the Gulf of Mexico dataset as the training set. create_rand_training_data.m performs a stratified 80/20 train/test split.

Parameter tuning

All files that begin with tune are for parameter tuning. tune_*_first_day_*.m are for the "first day" training set, and the others are for the 80/20 split.

Classifier training

train_*.m train classifiers with the tuned parameters.

Classifier testing and results

test_classifiers*.m test all trained classifiers on the relevant testing set and saves results. collect_cv_results.m collects cross-validation results obtained during parameter tuning.

Steps to reproduce results

In the files mentioned in the following steps, the path(s) to the data files need to be set. This variable is typically called box_dir. By default, this path is set to ../../data/fish-lidar/GulfOfMexico; therefore, you can avoid changing paths in all the scripts by putting the data archive in that directory.

Preprocessing

  1. Convert the PNG files into MAT files: script_extract_data_from_day.m
  2. Preprocess the data: create_processed_mat_files.m

Note: create_processed_mat_files.m uses a lot of RAM (>32 GB), so you either need more RAM than that or you need a fair amount of swap space.

First day results

  1. Create the training data: create_first_day_training_data.m
  2. Tune the undersampling ratio: tune_sampling_first_day_{lda,nnet,rusboost,svm.tree}.m
  3. Tune each model's hyperparameters: tune_hyperparams_first_day_{lda,nnet,rusboost,svm,tree}.m
  4. Tune the number of labels needed for a region of interest: tune_roi_first_day_{lda,nnet,rusboost,svm,tree}.m
  5. Train the final classifiers: train_first_day_{lda,nnet,rusboost,svm,tree}.m
  6. Collect cross validation results: collect_cv_results.m
  7. Test the classifiers: test_classifiers_first_day.m

80/20 split results

  1. Create the training data: create_rand_training_data.m
  2. Tune the undersampling ratio: tune_sampling_{lda,nnet,rusboost,svm.tree}.m
  3. Tune each model's hyperparameters: tune_hyperparams_{lda,nnet,rusboost,svm,tree}.m
  4. Tune the number of labels needed for a region of interest: tune_roi_{lda,nnet,rusboost,svm,tree}.m
  5. Train the final classifiers: train_{lda,nnet,rusboost,svm,tree}.m
  6. Collect cross validation results: collect_cv_results.m
  7. Test the classifiers: test_classifiers.m

Reproducing paper figures

  • Figures in the Datasets section of the paper can be produced by running example_data_figures.m
  • Figures in the Results section of the paper can be produced by running results_figures.m

You need to set the appropriate path to the data files in all of the scripts mentioned above.