Learning Must-Link Constraints for Video Segmentation Based on Spectral Clustering
read more
Citations
Going deeper with convolutions
Motion Trajectory Segmentation via Minimum Cost Multicuts
Evaluation of Super Voxel Methods for Early Video Processing (Author's Manuscript)
Guarantees for Spectral Clustering with Fairness Constraints
Submodular Trajectories for Better Motion Segmentation in Videos
References
Random Forests
Normalized cuts and image segmentation
Normalized cuts and image segmentation
A tutorial on spectral clustering
On Spectral Clustering: Analysis and an algorithm
Related Papers (5)
Frequently Asked Questions (17)
Q2. What is the main argument for spectral clustering?
Spectral clustering, as a relaxation of the NP-hard normalized cut problem, is suitable due to its ability to include long-range affinities [18, 40] and its global view on the problem [14], providing balanced solutions.
Q3. What is the problem of the normalized cut?
In particular, as the normalized cut itself is a NP-hard problem and even the spectral relaxation is non-convex, the optimization of the minimizer which yields the segmentation is out of reach.
Q4. What is the first set of experiments?
In the first set of experiments, the authors consider the Berkeley Motion Segmentation Dataset (BMDS) [8], which consists of 26 VGA-quality video sequences, representing mainly humans and cars, which the authors arrange into training, validation and test sets (6+4+16).
Q5. How has must-link constraint learning been tried?
The goal of integrating must-link constraints into spectral clustering has been tried via: i. modifying the value of affinities (cf. [24], which first considered constrained spectral clustering); ii. modifying the spectral embedding [30]; or iii. adding constraints in a post-processing step [49, 13, 48, 45, 33].
Q6. What are the main limitations of spectral techniques?
In this paper, the authors focus on two important limitations of spectral techniques: the excessive resource requirements and the lack of exploiting available training data.
Q7. How do the authors improve the performance of the learning algorithms?
the authors have shown that learned mustlink constraints improve efficiency and, in most cases, performance, as these allow discriminatively training on GT data.
Q8. What is the way to reduce the graph size?
Reducing the original graph size with learned must-link constraints allows to experiment with 1-SC on state-of-the-art video segmentation benchmarks [8, 17], notwithstanding the increased computational costs.
Q9. What is the affinity of the must-link constraining problem?
Must-link constraints have a transitive nature:Mpw(eij) = 1 andMpw(eik) = 1 imply Mpw(ejk) = 1. It is therefore crucial that all decided constraints are correct, as a few wrong ones may result in a larger set of incorrect decisions by transitive closure and potentially spoil the segmentation.
Q10. What is the function for a normalized cut?
Given a partition of V into N sets S1, . . . , SN , the normalized cut (NCut) is defined [31] as:NCut(S1, . . . , SN ) = N∑ k=1 cut(Sk,V\\Sk) vol(Sk) , (2)where cut(Sk,V\\Sk) = ∑ i∈Sk,j∈V\\Sk wij and vol(Sk) = ∑ i∈Sk,j∈V wij .
Q11. What is the description of the proposed method?
Thus learned must-links closely follow the spectral clustering optimization and their proposed method only provides further reduction of the problem size.
Q12. How does the proposed method reduce runtime?
With respect to the efficient reduction of [16], the authors further reduce runtime by 30% and memory load by 65%, while the authors reduce runtime by 97% and memory load by 87% wrt [18].
Q13. What is the reason why minimization of the NCut is NP-Hard?
The balancing factor prevents trivial solutions and is ideal when unary terms cannot be defined, but is also the reason why minimization of the NCut is NP-Hard.
Q14. What is the effect of learning mustlink constraints on the learning of a tree?
While this theory is applicable to general clustering and segmentation problems, the authors have particularly shown the use of learned must-link constraints in conjunction with spectral techniques, whereby recent theoretical advances employ these to reduce the original problem size, hence the runtime and memory requirements.
Q15. What is the way to learn Mpw?
From an implementation viewpoint, it is convenient to consider instead Mpw, defined over the set of edges E of the graph G representing the video sequence:Mpw : E 7→ {0, 1} (5)Mpw casts the must-link constraining problem as a binary classification one, where a true output for an input edge eij means that i and j belong to the same point grouping, in the must-link constrained graph GM .
Q16. How many false positives are made on the unseen data?
Although this conservative classifier might imply that in the worst case, no must-link constraints are predicted, it turns out their classifier actually predicts for a large fraction of the edges to be linked and thus leads to a significant reduction in size, while making a few false positives on the unseen data (overall, 1 false positive per 242k true predictions).
Q17. What are the main metrics for BPR and VPR?
Besides the PR curves, the authors report aggregate performance for BPR and VPR: optimal dataset scale [ODS], optimal segmentation scale [OSS], average precision [AP].