Skip to content

Latest commit

 

History

History
145 lines (109 loc) · 4.38 KB

annotating.md

File metadata and controls

145 lines (109 loc) · 4.38 KB
Raw

Annotating & AI Training

Annotation guidelines & notes

  • Use distortion2 for real distortions, not filters or fading.
    • The old term distortion is used for both distortion and filters.
  • Use high-pass and fade for filtered segments.
  • Use vechile, wind, rain, human tags only if especially loud, when it might be better to check these segments manually.
  • Tag silence only if no other tags. In practice, I have used this also with other sounds except animals.
  • Faint and strong have not been used systematically. They can be considered as good examples of faint and strong sounds.
  • Some dogs have been tagged as mammals.
  • Local_choir / individual and migrants can be tagged in same annotation.
  • Annotations are often lacking dogs etc, if there are only birds. So cannot be used for learning absence of dogs.

Todo

Annotating Xeno-Canto Sounds

  • ALWAYS DO

    • Replace spaces with underscores in filenames rename 's/ /_/g' *
    • Create nonmodified wav's for i in .mp3; do ffmpeg -i "$i" "${i%.}.wav"; done
    • Check that filenames don't have non-ascii characters, because they will choke exiftool: LC_ALL=C find . -name '[! -~]'

  • MAYBE DO

    • Basic augmentation sets:
      • Skip 2 sec, volume 1.3 for i in .mp3; do ffmpeg -ss 2 -i "$i" -filter:a "volume=1.3" "${i%.}-ss2vol13.wav"; done
      • Skip 5 sec, volume 0.7 for i in .mp3; do ffmpeg -ss 5 -i "$i" -filter:a "volume=0.7" "${i%.}-ss5vol07.wav"; done

  • Annotate problem recordings

    • distortion2 when spectrogram is malformed
    • fade
    • high-pass

AI Training

  • Classification

    • remove: ignore, distortion2, faded, (high-pass?)
    • animal: migrant, migrant-low, wander, local_individual, local_choir, owl, bat, mammal, dog, other_animal, mystery
    • no-animal: the rest

  • TBD

    • Clip top 15% away? -> faster handling, high-freq noise does not bother. Might affect negatively in recognizing spikes?
    • Remove most bats? Or learn to give positives on bats also?
    • include dog, mammal & other animal as positives?
    • remove faints that also have bats, noise or loud things?

  • don't use loud vehicles (at least from ks training recording), expect that recorded is not near roads. But use loud planes.

  • Crop images EXAMPLE: chop 50 pixels from top

    ../_portable/magick example.png -gravity North -chop 0x50 result.png

Biases etc.

  • no birds when loud vehicles (training data from winter, difficulties hearing)
  • most birds from noordwijk, with seashore wave sounds
  • many birds during good migration -> multiple calls in each segment

Old notes

Primary

  • no species TRAIN NEGATIVE
  • uncertain, skip (= leave out from training)

Birds:

  • nocmig, one (Single or few sounds passes by - nfc's) TRAIN
  • nocmig, many (Multiple similar sounds passes by - Melnig, Clahye, Braleu...) TRAIN
  • wandering, one (Single or few sounds passes by, but not interpreted as nocmig - Scorus) TRAIN
  • wandering, many (rare case?) MAYBE TRAIN
  • stationary, one (Few distincg sounds - Turpil, Turmer, first singer...) - TRAIN
  • stationary, many (Continuous sounds - morning chorus) MAYBE TRAIN
  • owl MAYBE TRAIN

Mammals:

  • bat NO TRAIN
  • mammal MAYBE TRAIN
  • mystery animal MAYBE TRAIN

Other, specify in notes

Disturbance

  • single sound (drops, cracks, etc.)
  • rain
  • wind
  • rustle
  • human
  • plane
  • vehicle

Observations

  • Species
  • NFC count
  • Individual count estimate
  • Flock count estimate
  • Notes

Misc

25.1.2020: 4376 annotations 647 migrant's (14,8 %) 578 migrant-low's (13,2 %)

31.1.2020: 5587 annotations 694 migrant's (15,9 %) 725 migrant-low's (13,0 %)