Talk given at the 9th Indo_Pacific Fish Conference, Okinawa, June 27th 2013
fish4knowledge.eu
pdf available from http://www.cwi.nl/~lynda/talks/2013/Fish4Knowledge130627.pdf
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
Fish4Knowledge: large scale coral reef fish monitoring using undersea computer vision methods.
1. Fish4Knowledge
Large scale coral reef fish monitoring using
undersea computer vision methods
Lynda Hardman
CWI (Centrum Wiskunde&Informatica), Amsterdam, NL
Centre for Mathematics & Computer Science
2. THE FISH4KNOWLEDGE
PROJECT
1. Collecting Video Data
• Collecting videos
• Detecting fish
• Recognizing fish species
• Recognizing fish behaviors
• Evaluating video analysis accuracy
2. Exploring Video Data
• Exploring fish counts
• Investigating species composition
• Checking potential biases
2fish4knowledge.eu
3. THE FISH4KNOWLEDGE
PROJECT
1. Collecting Video Data
• Collecting videos
• Detecting fish
• Recognizing fish species
• Recognizing fish behaviors
• Evaluating video analysis accuracy
2. Exploring Video Data
• Exploring fish counts
• Investigating species composition
• Checking potential biases
3fish4knowledge.eu
4. COLLECTING VIDEO DATA
Multiple video streams: 9 Cameras, >20,000 hours of videos (Dec. 2012)
Terabyte-scale data platform
High-performance servers for data access
4fish4knowledge.eu
5. DETECTING FISH
Detection of fish in each frame
Description of fish contour, color, texture... (>30 features)
Tracking of single fish over several frames
Handling partial and total occlusions
Over 400 million fish detections
5fish4knowledge.eu
6. RECOGNIZING
FISH SPECIES
Detection of body parts (head, tail…)
Construction of a fish model for each species
15 species detected (96% of total number of fish)
7. RECOGNIZING
FISH BEHAVIORS
Detection of behaviors by analyzing trajectories
Relation to background objects (feeding, hiding…)
Relation to other fish (pairing, grouping, solitary…)
Fish Trajectories Description of scene’s background
8. THE FISH4KNOWLEDGE
PROJECT
1. Collecting Video Data
• Collecting videos
• Detecting fish
• Recognizing fish species
• Recognizing fish behaviors
• Evaluating video analysis accuracy
2. Exploring Video Data
• Exploring fish counts
• Investigating species composition
• Checking potential biases
8fish4knowledge.eu
10. EVALUATING VIDEO
ANALYSIS ACCURACY
Certainty Scores indicate the quality of each detection
These can be used to filter out low-quality fish
Score > 0.8
Fish Counts for 5 Certainty Score Thresholds
Score > 0.2
Score > 0.4
Score > 0.5
Score > 0.6
Certainty Score: 0.4
Certainty Score: 0.9Certainty Score: 0.75
Certainty Score: 0.6
11. THE FISH4KNOWLEDGE
PROJECT
1. Collecting Video Data
• Collecting videos
• Detecting fish
• Recognizing fish species
• Recognizing fish behaviors
• Evaluating video analysis accuracy
2. Exploring Video Data
• Exploring fish counts
• Investigating species composition
• Checking potential biases
11fish4knowledge.eu
14. THE FISH4KNOWLEDGE
PROJECT
1. Collecting Video Data
• Collecting videos
• Detecting fish
• Recognizing fish species
• Recognizing fish behaviors
• Evaluating video analysis accuracy
2. Exploring Video Data
• Exploring fish counts
• Investigating species composition
• Checking potential biases
14fish4knowledge.eu
18. THE FISH4KNOWLEDGE
PROJECT
1. Collecting Video Data
• Collecting videos
• Detecting fish
• Recognizing fish species
• Recognizing fish behaviors
• Evaluating video analysis accuracy
2. Exploring Video Data
• Exploring fish counts
• Investigating species composition
• Checking potential biases
18fish4knowledge.eu
19. SPECIES COMPOSITION
DEPENDS ON LOCATION
Plectroglyphidodondickii, Camera 38 Amphiprionclarkii, Camera 38
Plectroglyphidodondickii, Camera 39 Amphiprionclarkii, Camera 39
Good afternoon! As you may have noticed, I am not Bas Boom, but Lynda Hardman, and I’m not a marine biologist, I’m a computer scientist.I am at this conference, because I am a member of a European project that has spent the last 3 years building a system for automatically counting fish in fish videos.
Here you can see the habitats for camera 38, on the left, and camera 39, on the right (also based near Kenting).