This repository contains Sloth annotations for extracting whale heads from whale images in the NOAA Right Whale Recognition Kaggle competition.
The initial dataset had 300 images over 11 whales:
smerity@pegasus:~/Coding/whale_hunt$ find heads/ | grep "jpg" | cut -d '/' -f 2 | sort | uniq -c
24 whale_08017
29 whale_24458
22 whale_26288
32 whale_28892
24 whale_34656
30 whale_36851
43 whale_38681
22 whale_48813
22 whale_48966
28 whale_51195
24 whale_52749
The new dataset has grown substantially as it has been helped by many!
The data is assumed to be in the imgs folder, sorted by identified whale as provided by inversion's Python script to sort the whale images.
To extract the heads, simply run:
python extract_tiles_from_sloth.py annotations/*.json
By default this will provide a 256 x 256 image of the head, expanded to be a box. If you woud like to maintain the exact annotation, set PAD_TO_BOX to False.
Whilst no explicit annotation guidelines have yet been formed, the head should capture all of the visible callosity patterns on the whale's head and include the eyes and eyebrow callosities if visible, even if obscured as they're underwater. If in doubt, more is better than less.
If you'd like to provide additional annotation, or augment the annotation yourself, you can do so by using Sloth, a versatile tool for various labeling tasks in the context of computer vision research.
find imgs/*/* -iname "*.jpg" | shuf | xargs sloth appendfiles annotations/whale_faces_smerity.json
ln -s annotations/whale_faces_smerity.json .
sloth --config slothwhales.py whale_faces_smerity.json
Replacing smerity with your identifying handle for easy merging. This will allow for easy merging of multiple annotations.
- inversion
- jfkingiii
- sunilsu
MIT License, as per LICENSE