detection identification - machine learningcs229.stanford.edu/proj2015/180_poster.pdf · • 4-fold...
Post on 29-Sep-2020
2 Views
Preview:
TRANSCRIPT
• K means local minimao Runs k means moreo Learn ocean and
whale centroids and initialize to those
• Better water removal
• 4-fold CV• 1 fold with at least 1 image
for each whale
Testing
Introduction Detection and Identification Algorithms
Dataset Results
The North Atlantic Right Whale is a critically endangered species that lives alongthe east coast of the United States. There are only an estimated 500 whales left inthe ocean and to aid conservation efforts, they must be tracked and cataloguedfrom aerial imagery [1,2].
Future Work
Credit: Woods Hole Oceanographic Institution Graphics
The task of identifying individual whales isusually performed by highly experiencedresearchers and marine biologists under stricttime and budget constraints.
Currently, the manual matching is conductedwith the aid of an online catalog and somecustom software [3] which is extremely timeconsuming.
[1] Kaggle Inc. Right Whale Recognition,https://www.kaggle.com/c/noaa-right-whale-recognition[2] NOAA Fisheries. North Atlantic Right Whales (eubalaena glacialis),http://www.nmfs.noaa.gov/pr/species/mammals/whales/north-atlantic-right-whale.html[3] DIGITS: Digital Image Gathering and Information Tracking System,http://www.neaq.org/conservation_and_research/projects/endangered_species_habitats/right_whale_research/right_whale_projects/monitoring_individuals_and_family_trees/identifying_with_photographs/digits.php
Raw Images
• Identify ROI• Save negative regions• Train detector
Cascade Object Detector
• Get 50 largest bounding boxes• Convert to mask
Whale Extraction• Dilate• Erode• Take largest blob
Mask Manipulation
• Detect SIFT features [4]• Apply PCA [5]
SIFT Extraction
• Convert RGB to HSV space• Random sample for cluster
centroids
Color Space Change
• K-means clustering, k = 4• Largest cluster is ocean• Invert to get whale mask
K-means Clustering
[4] VLFeat Team. (2013). VLFeat MATLAB Toolbox
http://www.vlfeat.org/index.html
This work is based on a Kaggle competition [1] that aims touse machine learning techniques to identify Right Whalesin a large dataset of aerial imagery.
# of images 11,000
# of training images 4,500
# of unique whales 447
Image resolution 4 – 17 MP
Poor lighting and
lack of contrast
from reflections on
water
Occlusions from
the sea foam
Distortion from the
water
Varying actions
and orientations
(breaching,
feeding)
• Pick whale with largest match %
SIFT Matching
• Remove uncommon features between images
Prune Features
• SVM feature is a single SIFT feature• Each SIFT feature is labeled for a
given whale• Entire image is then labeled as most
common label for SIFT features
SVM
Detection Identification
Detection
Identification
• SIFT comparison improvements to reduce false positives
• Continue working on SVM• Use other learning methods (e.g.
Convolutional Neural Networks)• Better SIFT features, or more
localized to certain parts of the whale (e.g. “face”, back, tail)
Ocean mask
K-mean mask
K-mean result
COD mask
COD result[5] Farag, A. A., & Elhabian, S. (2005). A Tutorial on Principal Component Analysis
http://dai.fmph.uniba.sk/courses/ml/sl/PCA.pdf
Color Space Detection Identification
ComputationTime
DetectionStep
Time(s/img)
K means ~60
COD ~10
SIFT ~5
IdentificationSteps
Time(s/img)
Pruning ~1
SIFT Matching ~80
SVM Testing ~500
Same whale:
Wrong whale:
Whale Detection and Identification from Aerial ImageryAditya Mahajan, Adrien Perkins
Department of Aeronautics and Astronautics, Stanford University
58 matches / 168 features
53 matches / 6654 features
Bad lighting will still show the whale in HSV space
HSV space candifferentiate whale, water and foam
Deep parts (like the tail)are also more apparent in HSV space
COD K-means
COD K-means
COD K-means
top related