R/fun_gtd.R

In georoen/wuepix: Using Computer Vision to Count Pedestrians

#' @author Jeroen Staab
# Ground Truth Data Tools
# Function
GTD_single <- function(img) {
  #' @title Sample Ground Truth Data
  #' @description Manually asses number of persons in a single image.
  #' @param img file path to image, also known as `now`.
  #' @return numeric vector with number of persons.
  #' @import jpeg
  #' @import magick
  #' @export GTD_single
  plot(as.raster(magick::image_read(img)))
  title(main = basename(img))
  # JPEG_plot(jpeg::readJPEG(img), basename(img))
  res <- readline("Please enter number of persons: ")
  if (res == "") {
    cat("EMPTY equals 0 persons.")
    res <- 0
  }
  res <- as.numeric(res)
  if (is.na(res)) {
    cat("Not numeric - try again!")
    res <- GTD_single(img)
  }
  res
}


GTD_list <- function(img.list) {
  #' @title Sample Ground Truth Data
  #' @description Manually asses number of persons in multiple images.
  #' @param img.list file path to image, also known as `now`.
  #' @return numeric vector with number of persons.
  #' @export GTD_list

  # Allocate output
  rtn <- img.list
  rtn[] <- NA

  # Loop
  p <- round(seq(1, length(img.list), length.out = 10)) # Progress
  for (i in 1:length(img.list)) {
    img <- img.list[i]
    rtn[i] <- GTD_single(img)

    if (i %in% p) {
      message(paste("Progress:", round(i / length(img.list) * 100), "%"))
    }
  }

  as.numeric(rtn)
}


# Accuracy Assesment
GTD_truePositives <- function(GTD, PRD) {
  #' @title Benchmark Pedestrian Detection
  #' @description Accuracy Assesment for Object-Based Classifiers as in DALAL
  #' etal. 2005 p. 888
  #' @return Returns datafram with following colums: FalseNegative,
  #' TruePositives, FalsePositive, MissRate, FalsePositvesPerWindow, correlation
  #' @param GTD Numeric vector of Ground-Truth-Data, as returned by GTD_list()
  #' @param PRD Numeric vector of prediction values, as returned by hog_list()
  #' and yolo_list()
  #' @export GTD_truePositives

  print(cor.test(GTD, PRD))

  df <- data.frame(GTD = GTD, PRD = PRD)

  df <- df %>%
    mutate(
      FP = PRD - GTD,
      FP = ifelse(FP >= 0, FP, 0), # positive values only
      TP = PRD - FP,
      FN = GTD - PRD,
      FN = ifelse(FN >= 0, FN, 0)
    ) %>%
    summarise(
      FN = sum(FN),
      TP = sum(TP),
      FP = sum(FP),
      MR = FN / (TP + FN),
      FPPW = FP / n()
    )
  df$cor <- cor(GTD, PRD)
  df
}