Wavesflag / toolbox.badacost.kitti.public

Multi-view car detection trained on KITTI dataset.

This repo has the auxiliary Matlab and C++ code in order to replicate the car detection experiments in our paper. If you use this code for your own research, you must reference our journal paper:

• BAdaCost: Multi-class Boosting with Costs. Antonio Fernández-Baldera, José M. Buenaposada, and Luis Baumela. Pattern Recognition, Elsevier. In press, 2018. DOI:10.1016/j.patcog.2018.02.022

Requirements

• Clone toolbox.badacost.public repo, with our modified version of Piotr Dollar toolbox with the BAdaCost algorithm with cost-sensitive trees. Go to its directory and execute Matlab. Then from Matlab prompt, execute addpath(createpath(PATH_TO_TOOLBOX)) and then toolboxCompile.

• From the object detection part of the KITTI database download:

  • The training images.
  • The training image labels.

  by decompressing the images file data_object_image_2.zip and the labels file data_object_label_2.zip we will get the following directory structure:

     kitti_database
      |
      +---training
      |      |
      |      +--- image_2 (png files for training)
      |      |
      |      +--- label_2 (txt files with ground truth)
      |             
      +---testing
             |
             +--- image_2 (png files fir testing and upload results to KITTI server)             
  

  The path, kitti_database in the example, with the KITTI training dir (with images and labels) will be refered as KITTI_PATH from now on.

Execution of the training scripts

There are two important scripts in the root of toolbox.badacost.kitti repository:

• main.m allows to prepare training data from KITTI_PATH, train a car detector with the best params and then test it on a set of a driving car images taken from KITTI server. Important variables to set in this script are:

  • TOOLBOX_BADACOST_PATH, path to the toolbox.badacost.public Matlab toolbox.
  • KITTI_PATH, path to the KITTI dataset.
  • PREPARE_DATA, set it to 1 to prepare KITTI data for BAdaCost training and execute main.m. Once prepared first time, you can set it to 0.
  • DO_TRAINING, set it to 1 to train the best parameters BAdaCost detector. Once trained first time, you can set it to 0.
  • FAST_DETECTION, set it to 1 in order to make faster detection but with less accuracy. Set it to 0 when you want improved accuracy as the cost of more execution time.
  • SAVE_RESULTS, set it to 1 in order to save processed images to disk (in the path given by IMG_RESULTS_PATH).
  • NICE_VISUALIZATION_SCORE_THRESHOLD, set it to the score value above detections are shown in results.
  • VIDEO_FILES_PATH, FIRST_IMAGE, IMG_EXT, are variables to point to the images over to execute the trained detector.

• main_paper_experiments.m allows to train SAMME and BAdaCost detectors with different parameters in bach.

• main_paper_experiments_SubCat.m allows to train SubCat detectors with different parameters in bach.

• main_kitti_test_best_detector.m allows to test the best SubCat, SAMME or BAdaCost best detector over the KITTI testing images.