Affine Motion-compensated Prediction

1
Affine Motion-compensated Prediction

• Dragomir Anguelov
• EE368B Final Project
• December 8, 2000

2
Motivation

• Motion models used for prediction
• Translation
• How about more complex motions rotation, shear?
• Tradeoff increased expressivity vs. increased
  number of parameters

3
Motion Models

• Translation 2 parameters
• Affine additional 4 parameters
• Account for rotation and shear

4
The Motion-compensated Hybrid Coder

• Testbed assumptions
• Want very good image reconstruction
• Details
• Subpixel accuracy
• No filtering
• DCT Entropy-Minimizing coder with uniform
  quantization
• Error image encoded and transmitted

5
The Hybrid MCHC
6
Affine Motion Estimation Overview

• Blockmatching techniques
• Infeasible, search space in 6 dimensions,
  optimizations in this space not well studied
• Differential techniques
• Lucas-Kanade pyramidal tracker (used in this
  project) Lucas, Kanade 81 Shi, Tomasi, 94
• Mixture techniques

7
Lucas-Kanade tracker

• Minimizes
• Newton-Raphson minimization using the derivatives
  of the error function
• Issues
• Assumes motions between frames are not too large
• Hierarchical implementation

8
Hybrid affine tracker

• Use blockmatching to determine several minima of
  the SSD function
• Initialize a differential method with those
  points and pick the best resulting set of affine
  parameters

9
Experimental results (1)

• Setup
• Padded image borders
• Affine parameter quantization with step 0.2
• DCT error quantization with step 1.0
• Block size 16

10
Experimental Results(2)

• Small, translation motion between frames
• Hybrid MCHC rates are comparable

Frame 1 2 3 4 Trate 4.1359 2.4845 2.4557
2.5346 Hrate 4.1359 2.4846 2.4556 2.5350 Diff
0 -0.0001 0.0001 -0.0004
11
Experimental Results (3)

• Relatively large motion, rotation present

Frame 1 2 3 4 TRate
5.6845 3.8926 3.8737
3.8393 Hrate 5.6845 3.8927 3.8702
3.8382 RateDiff 0 -0.0001
0.0035 0.0011
12
Experimental Results (4)

• Block Choices
• 0 translation 1 affine

13
Experimental Results (5)

• Block size 32, param quantization step 0.1
• Frame 1 2 3 4
• Trate 5.8776 4.0643 4.0980
  4.0747
• Hrate 5.8776 4.0608 4.0851
  4.0704
• Diff 0 0.0035 0.0131 0.0043

14
Conclusions

• Not too effective system
• Small improvements
• Increases running time
• Limitation on accuracy of transmission and amount
  of motion
• Improvement (better rate, same quality)
• Larger blocksizes
• Complex motions