Scalable Feature Extraction for Remote Motion Estimation

1
Scalable Feature Extraction for Remote Motion
Estimation

• Yao-Chung Lin
• Yun-Hsuan Sung
• Shu-Ping Yeh

2
Outline

• Introduction
• N-Queen Spatial Decimation
• DCT based Hash Code
• Simulation Results
• Conclusion and Discussion

3
Introduction

• Application
• Low complexity video encoder
• Wireless video sensors for surveillance
• Mobile camera phone
• Approach
• Dual of motion estimation at encoder
• Spatial decimation
• Frequency selection

4
Remote Estimation Model
X Current Frame to be coded Y Side information
(previous frame) Z Reconstructed frame from T
Y T Hash Code C Compressed Hash Code
5
Outline

• Introduction
• N-Queen Spatial Decimation
• DCT based Hash Code
• Simulation Results
• Conclusion and Discussion

6
N-queen decimation

• Divide one frame into 4 x 4 or 8 x 8 blocks
• Sub-sample by 4-Queen or 8-Queen Pattern

41 sampling rate
81 sampling rate
7
Macro block

• Decimation increases the ambiguity between
  blocks.
• Using Macro block as searching basic unit.

4-Queen
8-Queen
8
2-D DCT for 4-Queen

• Rearrange 4-Queen into 2 x 2 block
• Form 8 x 8 macro block and apply 2-D DCT
• JPEG-LS Predictor for DC terms
• Uniform quantization for AC terms

9
1-D DCT for 8-Queen

• Rearrange the 8-Queen into 1-D vector.
• Apply 1-D DCT
• Use JEPG-LS Predictor for DC coefficient
• Use different step size for each AC coefficients

10
DCT decimation

• Decimate the DCT coefficients one-by-one from
  high frequency terms
• Provide scalability

11
Outline

• Introduction
• N-Queen Spatial Decimation
• Adaptive DCT Coefficient Selection
• Simulation Results
• Conclusion and Discussion

12
DCT Based Hash Code

• DCT Coefficient distribution characteristics
• DC Uniform distribution
• AC Laplacian distribution

13
Generating Hash Code

• Parameter
• b_DC bits used to code DC coefficient
• nCoeff number of coefficients to be coded.
• DC
• AC

14
Quantities of Classified Coefficient
15
Compression

• DC
• Down sampled picture
• JPEG-LS predictor
• AC
• Noise
• No prediction Used

16
Block Matching Criteria

• Match the block in hash code domain
• Sum of estimated MSE of available coefficients
• Select the minimum estimated MSE

17
Collision

• r bits for coding the block
• (M1)2 blocks in search window
• What happened if ?
• 2-Stage Block Matching
• Larger block, less collision
• Refine each block in displacements,
  clip(Mr,0,10)
• Where

18
Two-stage Search

• Large Block Search Combine 4 macroblocks/blocks
  to do the first stage search.
• Refinement Search Search the matching block in
  the previously found window.

19
Outline

• Introduction
• N-Queen Spatial Decimation
• DCT based Hash Code
• Simulation Results
• Conclusion and Discussion

20
Simulation Setting

• Estimate rate by entropy
• Compute distortion between original and
  compensated
• Compare improvements
• Compare two methods
• Code first 50 CIF frames

21
Results of N-Queen
22
Results of Adaptive DCT Coefficient Selection
23
Dog Fighting
24
Summary of Simulation Result

• Significant improvement of two-stage search
• Reduction by coefficients provides scalability
• Two methods have almost the same performance

25
Conclusion

• Proposed
• Spatial domain approach
• Frequency domain approach
• Sophisticated motion estimation
• Future Work
• Reducing collision
• Reserving feature information
• Optimizing bit allocation