Reads validation_by_clip.csv, the clip-by-clip result of the review
step of the 'sheepdog' package, and returns the verdicts in a form that
fp_aggregate() can attach to the occasions. Files written by the
sampler up to version 1.2.0
(validacao_por_clipe.csv, with column names in Portuguese) are read as
well.
Value
A data frame with one row per clip: site, species_code,
file (the audio file the clip was cut from), start and end (in
seconds within it), confidence, verdict (TRUE for a true
detection, FALSE for a false one, NA for clips not reviewed), and
occasion_id when the file has it.
Details
The format is detected from the columns: site marks the format of
'sheepdog' 2.0.0 and later, local the legacy one. The columns used are
site, species_code, source_audio, clip_start, clip_end,
confidence, and verdict (legacy: local, species_code,
audio_origem, inicio_clipe, fim_clipe, confidence, veredito),
and occasion_id (ocasiao_id) when present. Verdicts positive and
negative (legacy: positivo, negativo) become TRUE and FALSE;
not_reviewed and clip_missing (nao_validado, clipe_ausente)
become NA, and any other value becomes NA with a warning.
The sampler can pick clips among the detections of a species by their
confidence or at random. The models in crywolf treat the checked
occasions as chosen independently of the classifier's score: clips
picked for their high scores hold fewer false positives than a random
sample, and would make the false-positive rate look lower than it is. Use
clips drawn at random among the detections for this purpose. Clips of
whole recordings (sampling_mode = "random_recordings", with no
confidence) are a second, manual method of detection: give them to
fp_data() as verified occasions, not to the validation argument of
fp_aggregate().