Name: Common Banded Peacock Scientific Name: Papilio crino Species ID #: 12324343553354 Family: Papilionidae Genus: Papilio Distribution: Central and Southern India, and Srilanka Status: Locally Common Wing Span: 100 – 116mm Conservation Status: Not threatened ID-Me : Tool for identifying species from an image Kishen Das Kondabagilu Rajanna UT Arlington, Texas
The tool is in its initial stages. It has to yet tested against a huge repository, say around 300 species. If you have any suggestions please mail to [email protected]
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.
SIFT in a nutshell•Keypoint detection and localization
( Gaussian Filtering)
•Orientation Assignment
Peak Thresh=1.0
Edge Thresh=2000
Match Thresh=1.5
Keypoint ->
4 X 1 matrix
(Scale,
Orientation,
Translation along x,
Translation along y)
Descriptor ->
128 X 1 matrix
•Orientation Assignment
•Keypoint descriptor matching
( Nearest neighbough indexing)
1. Extract descriptors for dt and dq
2. Extract first closest descriptor d1
3. Extract second closest descriptor d2
4. Accept d1, if dist(dq,d2) > dist(dq,, d1)
Edge threshold eliminates peaks of the DoG scale space whose curvature is too small
Peak threshold filters peaks of the DoG scale space that are too small (in absolute value)Peak threshold filters peaks of the DoG scale space that are too small (in absolute value)
Credits: http://www.vlfeat.org/overview/sift.html
199
81
Matches
51
Matches
199
Matches
Hough Transform in 4D space
Step 1) Prepare the bins based on David Lowe [3]
Step 2) Each matching votes for 4 bins ( In David Lowe [3],each matching votes for 16 bins, to avoid the boundary effects. I am considering only 4 bins for simplicity)
Step 3) Find out the configuration [ Scale Bin with maximum votes, Orientation Bin Step 3) Find out the configuration [ Scale Bin with maximum votes, Orientation Bin with maximum votes, X Translation Bin with maximum votes, Y Translation Bin with maximum votes]
Step 4) Choose the cluster of keypoints that has voted for the above configuration and discard rest of the clusters.( Ideally you should consider all the other clusters as well before discarding the keypoints, again for simplicity I am considering only one configuration with maximum votes) .
25
10
Matches
6
Matches
25
Matches
Modifed RANSAC in 4D space for
Affine TransformationStep 1) Group the remaining matches into all possible combinations of groups of 3
Step 2) For each group of 3-matches, find the differences of affine transformations between each member of this group and rest of the matches. Check whether the differences are within the margin of "OriginalBinSize/3". If yes, that's an inlier.yes, that's an inlier.
Step 3) Select group of 3-matches, if there are at least 10 inliers wrt that group.
Step 4) From the groups of 3-matches selected in previous step, pick the final individual matches, such that more than 50% of the groups have been picked in step 3 where this match belongs to.
Modified Homography for
uncalibrated camers
Step 1 ) Take 4 matches , estimate Homography matrix using the equation
∑ [ x training_image]X H [ x_query_image] < t, where t is close to zero.
If 't' cannot be closer to zero, then those set of matches don't belong to the same plane and hence discard them.
Here the above equation is converted into the form Ax = O ( Zero vector) and then Homography matrix is estimated using SVD.Here the above equation is converted into the form Ax = O ( Zero vector) and then Homography matrix is estimated using SVD.
Step 2) Keep repeating Step 1) till there are maximum number of inliers such that
∑ [ x training_image]X H [ x_query_image] < t, where t -> 0
Step 3) Discard outliers that will not fit into Homography constraint.
0
Matches
23
Matches
0
Matches
Matches
Related Work
http://www.ifpindia.org/biotik/index.php
In this tool, one can try to identify trees of ever green forest by building a query based on different tree parts (Leaf, branch, flower, etc
http://ippcweb.science.oregonstate.edu/LepID/
This software comes somewhat close to ID-Me. The major drawback is that the user has to know which part of the wing is important in identifying that species.
In this approach tool will automatically extract the wing venation and later identification will be done using 2 different types of artificial neural networks , multi-layer perceptron and learning vector quantisation
http://ipmnet.org/bugwing/
Its a simple tool to assist the amateur ecologists with identification. Here veins and their basic sequence of branching have been made use of in distinguishing the insects to the family or subfamily level.
Future Work
•Using Color descriptors
•Removing current limitations
•Using database to store keypoints and descriptors
•More images for each Training Image
•Fine tuning existing algorithms
•Introducing confidence levels
•Collaborations with taxonomists and computer scientists