photo matching

Document Sample
photo matching Powered By Docstoc
					 Photo-based identification
for Central Texas Eurycea: a
 viable substitute for VIEs?
          Nathan F. Bendik
Background
• Photography based ID is nothing new
• Many studies have relied on manual, “by-eye” matching
  • E.g. COA has been using photo-ID on Eurycea for many years to
    identify individuals in captivity
• BUT Manual matching is impractical for large problems
Background
• Photography based ID is nothing new
• Many studies have relied on manual, “by-eye” matching
  • E.g. COA has been using photo-ID on Eurycea for many years to
    identify individuals in captivity
• BUT Manual matching is impractical for large problems
• Computer based solutions to matching
  • Custom solutions for large (usually well funded) projects
  • Free software
  • Interactive Individual Identification System (I3S) Manta
     • Uses spot pattern recognition- not really suitable for Eurycea
  • Custom Algorithm for marbled salamanders: Gamble, Ravela and
    McGarhigal 2008
Wild-ID (Bolger et al. in press)
• Open source (free and free)
• Java-based (e.g. can use on Windows, Mac, Linux)
• Does not require tracing spots, patterns, etc.
Wild-ID (Bolger et al. in press)
• Open source (free and free)
• Java-based (e.g. can use on Windows, Mac, Linux)
• Does not require tracing spots, patterns, etc.

• Pattern matching and extraction using SIFT: Scale Invariant
  Feature Transform (Lowe 2004)
• Extracts image features invariant to image scale and rotation
• SIFT features are extracted and compared for each pair of
  photos- then a score is generated based on goodness of fit
• Scores are used to rank matches
• The user compares (up to) the top 20 ranked images and
  selects the best match by eye
Wild-ID




          Bolger et al. 2011
Wild-ID
•
How well does matching work?
  • False rejection: Does Wild-ID correctly match known unique
    individuals?
     • Does it rank true matches highest?
  • False acceptance: Does it incorrectly match different individuals?
     • Does it assign a high score to non-matches?


• Test subjects
  • Captive matched-by-eye E. sosorum photos (all recaptured
    individuals)
  • Wild E. sosorum photos
  • Wild, VIE-marked Eurycea tonkawae (many recaptured and
    single-captured individuals)
Potential problems
• Poor quality photographs:
  •   Low resolution photography
  •   Out of focus
  •   Poor exposure
  •   Poor contrast (poor lighting conditions)
  •   Motion blur (slow shutter speed/incorrect flash settings)
  •   Distortion (crappy camera lens)
  •   Different angles (e.g. head tilted, subject to focal plane not
      parallel)
Potential problems
• Poor quality photographs:
  •   Low resolution photography
  •   Out of focus
  •   Poor exposure
  •   Poor contrast (poor lighting conditions)
  •   Motion blur (slow shutter speed/incorrect flash settings)
  •   Distortion (crappy camera lens)
  •   Different angles (e.g. head tilted, subject to focal plane not
      parallel)
• “Evolving natural marks” (Yoshizaki et al. 2009)
  • Injury (e.g. position and number of chromatophores)
  • Growth and development (e.g. density of chromatophores)
  • Environmental (e.g. size of chromatophores)
Captive E. sosorum
•   Known matches from manual photo ID
•   Photos generally poor quality
•   Some examples of evolving natural marks
•   26 unique individuals (originally matched by eye)
•   2-6 photos per individual (85 total photos)
•   3570 photo pairs scored
•   104 possible correct matches
Captive E. sosorum
• Known matches from manual photo ID
• Photos generally poor quality
• Some examples of evolving natural marks
• 26 unique individuals (originally matched by eye)
• 2-6 photos per individual (85 total photos)
• 3570 photo pairs scored
• 104 possible correct matches
• Only considered highest ranked photos as matches- all others
  considered error
• Assumption: same individual will not change to look like other
  individual, and vice versa
Results: captive E. sosorum
•   9/104 did not score true match as highest
•   9% false rejection rate
•   0/85 matches to wrong individual
•   0% false acceptance
•   Many really bad photos
•   Good photos = high score (>0.1) when there is a real match
Wild E. sosorum
•   Captured on same day
•   35 total individuals
•   7 matched pairs (same individual, different photo angle etc.)
•   29 singles
•   Poor to very poor photo quality
•   Evolving natural mark errors not possible
Results: Wild E. sosorum
•   1/7 did not score true match as highest (blurry, small photo)
•   0/35 matched incorrect individual
•   Good photos = high score (>0.1) when there is a real match
•   Lower rejection rate with better pictures?
VIEs vs. Photos: E. tonkawae and
          score filtering
• Six 3-day capture sessions from 2008-2010
• 1350 photos, 725 unique individuals
• ~910 thousand photo pairs
• All were marked with VIEs and VIEs were confirmed by
  matching photos manually
• Goal = save time but maintain accuracy
    • Add more automation but maintain interaction
VIEs vs. Photos: E. tonkawae and
          score filtering
• Six 3-day capture sessions from 2008-2010
• 1350 photos, 725 unique individuals
• ~910 thousand photo pairs
• All were marked with VIEs and VIEs were confirmed by
  matching photos manually
• Goal = save time but maintain accuracy
    • Add more automation but maintain interaction
• Strategy:
    •   Assume rank 1 = potential match
    •   Evaluate score distribution (histogram)
    •   High score = match
    •   Low score = not a match
    •   Intermediate score = maybe -> visual check
Rank 1 scores for all matched pairs



=Not Match?




              =Match?




                  Score
Rank 1 scores for all matched pairs:
        Correct pairs (red)
      Incorrect pairs (green)




                   Score
VIEs vs. Photos
• Check 100 rank 1 pairs with score < 0.1
• Missed one real match
• But we left out potential non-rank 1 matches….
• Construct capture histories for each unique individual based
  on photo matching
• Difference in number of unique individuals indicates
  incongruence between VIE and photo IDs
VIEs vs. Photos: Results
• WildID + my score filtering scheme
  •   3/725 incorrectly ID’ed as unique individual
  •   False rejection rate <0.4%
  •   False acceptance rate of ZERO
  •   2 errors from changing natural marks
  •   1 error from poor photo quality
VIEs vs. Photos: Results
• WildID + my score filtering scheme
  •   3/725 incorrectly ID’ed as unique individual
  •   False rejection rate <0.4%
  •   False acceptance rate of ZERO
  •   2 errors from changing natural marks
  •   1 error from poor photo quality
• Compared to VIE?
  • Misidentification errors
  • Mis-mark errors (double marking, etc.)
  • 15/725 incorrectly ID’ed as unique individual
High score, but non VIE match.




             Score
Conclusions
• WildID photo matching >> VIE marking
  •   Lower error rate
  •   Less time required for data collection and data management
  •   Cheaper
  •   Less invasive
Conclusions
• WildID photo matching >> VIE marking
  •   Lower error rate
  •   Less time required for data collection and data management
  •   Cheaper
  •   Less invasive
• Manual checking some matches is required, but sorting scores
  and using photo database makes this process exceptionally
  fast
• Larger problems may require more interaction
• Using R and Lightroom can help scale problems to manageable
  sizes because high scores seem to guarantee a correct match
Sample Workflow
•   Photograph using DSLR and macro lens and close-up flash
•   Use standard non-busy background (e.g. white with gridlines)
•   Have camera name files sequentially (e.g. from 1 to 9999)
•   Append all photo names with date
•   Rotate and crop heads*
•   Export cropped images to folder (do not edit originals)
•   Run Wild-ID
•   Use R or SAS to convert output to capture history file
Acknowledgements
• Dee Ann            • Mark Sanders
  Chamberlain        • Blake Sissel
• Liza Colucci       • Bennett Vance
• Laurie Dries       • Many other temps,
• Andy Gluesenkamp     volunteers, field
• Todd Jackson         hands
• Tom Morrison       • LCRA
• Lisa O’Donnell     • TPWD
• Aaron Richter      • COA Water Utility
Captive E. tonkawae F1
                         1 month old    12.0 mm BL, 21.7 mm TL




                         6 months old   21.0 mm BL, 40.2 mm TL
  TL = 43.5 mm BL = 23.0 mm          TL = 24.1 mm BL = 13.3 mm


• Small animals that grow fast may pose the biggest problem
• BUT- can easily sort to focus manual matching effort on small
  individuals
                         Original




Gaussian blur radius=2              Gaussian blur radius=4




      Score: 0.8                          Score: 0.6
Individual 27: Incorrect match
Individual 27: Correct match
Correct match; small melanophores

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:21
posted:8/17/2012
language:
pages:36