Sparse representation for coarse and fine object recognition

1
Sparse representation for coarse and fine object
recognition

• Thang V. Pham Arnold W. M. Smeulders
• ISIS research group
• University of Amsterdam, the Netherlands

AIO-SOOS
2
Content

• Coarse and fine recognition with PCA
• A new representation
• Gaussian derivative bases
• Experimental results
• Conclusions

3
Coarse and fine recognition
Bear
90 degree
0 degree
Car
Name? Pose?
Duck
Training
Testing
4
with PCA
project
bear
project
eigenspace
car
duck
5
Our assessment of PCA-based

• Storage space is huge
• Recognition time is large
• Spatial coherent not exploited
• identical result by permuting the pixels
• Incremental learning absent
• large datasets
• Inefficient when unknown object localization

6
Our idea

• Sparsity
• Each object uses a small number of N bases from
  a potentially very large dictionary.
• Typically N could go up to 1000
• from a dictionary up to 2003.

7
Coarse and fine recognition
To model orientation, an image is modeled as a 3D
function
Each basis is separable
Each 1D basis is a Gaussian derivative
8
with local bases
Reconstruct
Compare
new object
duck space
bear space
9
Remember our idea

• Sparsity
• (Each object uses a small number of N bases from
  a potentially very large dictionary.)
• by matching pursuit
• Initialize residual object images
• Select the best basis for the current residual
• Update the residual
• Goto 2 unless the number of bases N

10
So far, in contrast to PCA

• Storage space is efficient
• No sampled points
• Not the axes, but their indices in the
  dictionary.
• Spatial coherence is exploited
• Yes, there is incremental learning
• - Some loss for recognition and localization?
• Is Reconstruct and compare inefficient?

11
The answer is NO.

• Approximating bases by piece-wise polynomials, we
  turn matching to polynomial evaluation.

12
Polynomial recognition

• A new image is recognized by
• Compute the piece-wise polynomials
• Compute the complete Njet coefficients of the
  test image at all locations
• Select N coefficients for each object from the
  object models as learned.
• Polynomial computation to yield polynomials (of
  degree 6 max).
• Evaluating the polynomials along the orientation
  to find the best matching candidate.

13
So far, in contrast to PCA

• Storage space is efficient
• No sampled points
• Not the axes, but their indices in the
  dictionary.
• Spatial coherence is exploited
• Yes, there is incremental learning
• Recognition phase is fast
• - / Njet is slower than PCA but done once
• Efficient for localization
• Efficient for many objects

14
Experimental results
15
Experimental results
16
Experimental results
1000 bases
6-D eigenspace
100-D eigenspace
17
Experimental results
18
Conclusions

• Efficient in storage space
• (no sample points)
• Efficient in recognition time
• (polynomial evaluation)
• Spatial correlation exploited
• (in framework of Gaussian differentials)
• Efficient object localization multiple objects
• (work with Njet coefficients)
• Large dataset and incremental learning
• (no re-training of existing models)