Skip to content
Machine learning final project
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
32_test.py
Niall Williams and Harry Zhou Whale Identification.pdf
README.md
Whales Poster.pdf
bad_model.py
base.py
count_distrib.py
filter_crops.py
generate_align_points_csv.py
log.txt
lr_model.py
match_photo_dirs.py
model.py
points1.json
points2.json
position_pic.py
position_pic_128.py
rotate.py
separate_images.py
train.csv
working_bad_cnn.py

README.md

Data source: https://www.kaggle.com/c/noaa-right-whale-recognition

Blog post explaining the winner's methodology: https://blog.deepsense.ai/deep-learning-right-whale-recognition-kaggle/

Paper on the techniques the winners used: https://arxiv.org/pdf/1312.6229.pdf

Head annotation training data: https://www.kaggle.com/c/noaa-right-whale-recognition/discussion/1742

Second place github with link to blog post: https://github.com/felixlaumon/kaggle-right-whale/tree/master/scripts

Keras: https://keras.io/

Really easy to understand article about using Keras for localization: https://towardsdatascience.com/object-detection-with-neural-networks-a4e2c46b4491

WHALE TRIG: https://github.com/anlthms/whale-2015/blob/master/object-recognition.pdf

CNN tutorials:

TO DO:

  • DETERMINE DISTIBUTION OF WHALES IN TRAIN SET

  • Train!!!

NOTES TO INCLUDE IN PAPER

  • We we unable to use the test dataset because the kaggle competition has stopped accepting entries. Instead, we just split our training data into train and test sets.

  • We lost 3 training images due to corruption. Ended with 416 different whales.

  • Original competition had 4 months. we had 3 weeks

  • Some more images lost when rotating anf cropping, because the images are awful quality

  • We did not have the hardware available to run 256 and 512 image scripts

  • Winner of the competition had an 87% accuracy...

  • Good images to compare: w_107.jpg shows how better crop removes the black edges from rotations. w_1112.jpg is a good indicator of the difficulty when there it a lot of sun reflection.

You can’t perform that action at this time.