1. Online Graph-Based Tracking
Postech Computer Vision Lab.
Hyeonseob Nam

2. Outline Introduction Algorithm Overview Main Algorithm Results
Density Propagation
Weighted Density Aggregation
Model Update
Results

3. Visual Tracking
Estimate the target state throughout an input video.

4. Motivation
Linear chain model assumes the temporal smoothness of two consecutive frames.
However, it usually breaks because of fast motion, shot change, occlusion, etc.
We propose an algorithm which constructs a graphical model considering the tracking appropriateness. 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

5. Traditional Linear Model
Sequentially estimate the target posteriors by the temporal order of frames. 1 29 34 43 50 92

6. Bayesian Model Averaging
Tracking easy-to-track frames first, estimate the target posteriors by blind model averaging. 1 29 34 43 50 92
Hong, S., Kwak, S., Han, B.: Orderless tracking through model-averaged posterior estimation. In: ICCV. (2013)

7. Our Approach
Keep the temporal tracking order, but select the relevant frames where the posteriors are propagated from. 1 29 34 43 50 92

8. Our Approach

9. Complexity Issue
Too many candidate!!
Current frame
Tracked frames

10. Complexity Issue
Current frame
Tracked frames
Representative frames

11. Algorithm Overview Density Propagation Density Aggregation
Propagate the density functions by patch matching
Density Aggregation
Aggregate the posteriors regarding the tracking plausibility
Model Update
Update the set of representative frames

12. Density Propagation Implement Bayesian filtering by patch matching
: A set of samples drawn from
Hong, S., Kwak, S., Han, B.: Orderless tracking through model-averaged posterior estimation. In: ICCV. (2013)

13. Density Propagation Matched patches Vote to center Tracked frame 𝒖
New frame 𝒕
Matched patches
Vote to center
Sample 1
Voting map by sample 1
Hong, S., Kwak, S., Han, B.: Orderless tracking through model-averaged posterior estimation. In: ICCV. (2013)

14. Density Propagation
Sample 2
Sample 3
Sample 4
Sample |𝕊 t |
Posterior 𝒕 …
votes 𝒕 𝒌
Hong, S., Kwak, S., Han, B.: Orderless tracking through model-averaged posterior estimation. In: ICCV. (2013)

15. Weighted Density Aggregation
Weighted Bayesian Model Averaging
Key problem: How to determine the posterior weights?
: A set of representative frames when tracking frame t

16. Determining the weights
𝛿 𝑢→𝑡 : The potential risk resulting from tracking 𝑢→𝑡
𝑑 𝑐 (𝑢,𝑡): Deformation cost by patch matching
𝛿 𝑢 : The minimax distance from the initial frame to 𝑢

17. Determining the weights
𝑑 𝑐 (𝑢,𝑡): Deformation cost by patch matching
Current frame: 𝑡
Target at frame 𝑢
Temporal target at frame 𝑡
𝜏 𝑢
𝜏 𝑡

18. Determining the weights
𝛿 𝑢 : The minimax distance from the initial frame to 𝑢
Current frame: 𝑡
Current graph structure
𝛿 1 =0 3 7
𝜹 𝟑→𝒕 =𝟒 8
𝛿 2 =3 10
𝜹 𝟐→𝒕 =𝟖 4 3 1
𝜹 𝟒→𝒕 =𝟏𝟎
𝛿 4 =7
𝛿 3 =4

19. Model Update Classification False positives Model Update
New target
Classification
False positives
Model Update
Redundant or useless templates
Representative frames

20. Template Classification
Prevent false tracking results from entering the set of representative frames
A set of positive and negative templates
New target template 𝜏 𝑡 ∗ is positive if the following measure is less than a threshold.
𝑆 𝑝 : The average Euclidean distance between 𝜏 𝑡 ∗ and k-nearest positive templates in 𝐷 𝑡−1
𝑆 𝑛 : The Euclidean distance between 𝜏 𝑡 ∗ and nearest negative templates in 𝐷 𝑡−1

21. Maintaining Representative Frames
Each template in the representative set should be distinct and useful for further tracking.

22. Maintaining Representative Frames
Distinctness:
Usefulness:
Template Weight:
Update:
where
where
where

23. Results

24. Identified graph structure

25. Quantitative Results
Benchmark
Challenges

26. Quantitative Results

27. Summary
We propose an adaptive and active algorithm to identify a general graphical model for a robust tracking.
A new target posterior is estimated by a selective and weighted model averaging.
For efficiency, only a small number of frames capture the important characteristics of input video.
Outstanding experimental results on 50 sequences in the tracking benchmark and 10 more challenging sequences show the benefit of our progressive graph construction algorithm.