Kaggle Deep Learning Competition in Computer Vision Alexander Eckert <[email protected]> Oct, 2018 Thanks to Nicole Finnie for preparing the bulk of the slides!
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
KaggleDeep Learning Competition
in Computer Vision
Alexander Eckert <[email protected]>
Oct, 2018
Thanks to Nicole Finnie for preparing the bulk of the slides!
Agenda
l A very brief introduction of Kaggle
l Kaggle 2018 Data Science Bowl – Nuclei detection
l Q & A
https://www.kaggle.com/c/data-science-bowl-2018
l Brief introduction of Kaggle
Prizes | Medals | Points
https://www.kaggle.com
Ranking & Rewards
l Competition ranking- Public leaderboard
l evaluated on subset of ground truthl visible during competition
- Private leaderboard l unseen datal determines final ranking
- Rewardsl Money, medals, ranking points
l Global ranking- Competitions- Kernels (Notebooks)
- Forum discussions
https://www.kaggle.com
Our team
https://www.kaggle.com/c/data-science-bowl-2018
l Our objectives- Apply and learn state of the art CV algorithms, ANN frameworks & architectures
- Take perspective of a data scientist, discuss approaches
- Compete
l 2018 Data Science Bowl – Nuclei Detection
Detect
Mask of one nucleushttps://www.kaggle.com/c/data-science-bowl-2018
Exploratory Data Analysis (EDA)K-Means
(Mini) U-Net – a CNN autoencoder
16 (3x3) filters
64 (3x3) filters
128 (3x3) filters
32 (3x3) filters
256 (3x3) filters
32 (3x3) filters
dropout 0.1
dropout 0.2
dropout 0.2
dropout 0.3
dropout 0.1
activation: Sigmoid
optimization: adamloss: binary cross entropy
input: 256x256x2 output: 256x256x2
https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/
Hidden feature engineering
KNN on edge pixels
channel 0 channel 1train image
windowing crops, flip(data augmentation)
2 U-Net modelscolour & grey
Train
Post processing of predictionsl Necessary for instance segmentation
l (Predicted masks – predicted contours) => label seeds
l Labelling (using random walker/watershed algorithm)
ground truth
post-processed
labelling (instance segmentation)predicted contours
test image
Model ensemblingbaseline
transform
noises
(data augmentation)
Pixel prediction using weighted majority vote
Final mean IoU: 0.545
ensembling
train images
test image
https://github.com/nicolefinnie/kaggle-dsb2018
Silver medal
Learnings
l Very time intensive- Important to efficiently (e.g. in parallel) evaluate promising paths and check classification results/errors
- GPUs and patience required for (deep learning) model training
l Follow forum discussions for new ideas, knowledge sharing and competition timeline
l Combine orthogonal approaches and create weighted ensembles
l Steep learning curve, but great for acquiring (or improving) skills- Keras/Tensorflow & PyTorch NN frameworks
- Python libraries like OpenCV, Pandas, Scikit, Jupyter
- Image processing techniques
- Network architectures for semantic segmentation
- Building modelling & prediction pipelines
Q & A
Sometimes it can go really wrong... (toxic image)What? A painting
in the test set?
Good job, U-Nets!
Related Documents
