Q2. What are the main tasks investigated for general robot control in real environment?
Obstacle detection and avoidance are the main tasks investigatedfor general robot control in real environment exploiting optical flow [57,67,82].
Q3. What is the basis of graph cuts?
The basis of graph cuts is the maxflow/min-cut problem consisting in finding the optimal path between two nodes in a directed graph, solvable by many algorithm in polynomial time [91,102].
Q4. What is the way to model smooth motion fields?
Low-order polynomials are usually sufficient to model smooth motion fields, and their small number of parameters allows for efficient computation.
Q5. What is the spectacular contribution of patch matching?
Research on speeding up block matching include multi-scale search strategies [243], integral images [83] or search in trees[158], but the recent most spectacular contribution was achieved in [15,16] with the PatchMatch algorithm.
Q6. What is the main undesirable effect produced by smoothing at coarse levels?
The main undesirable effect produced by smoothing at coarse levels is the loss of small and rapidly moving objects in the final estimated flow field.
Q7. What are the two common error measures used when ground truth is available?
When ground truth is available, two error measures are commonly used, namely the Angular Error (AE) and the Endpoint Error (EPE).
Q8. What is the optimal optimization of the data term with v fixed?
The optimization of the data term with v fixed is efficiently performed by a thresholding scheme [280], and fixing w yields the Rudin– Osher–Fatemi model [210], optimized with the duality based algorithm of [54].
Q9. What is the way to model an undirected MRF structure?
It is possible to model an undirected MRF structure as a directed graph by introducing two additional source and sink nodes, and then interpret the min-cut partition as a binary label segmentation of the MRF energy.
Q10. What is the main research direction for the large scale problem of optical flow?
To avoid the drawbacks of coarse-to-fine schemes of variational methods, an active research direction focuses on the design of computationally tractable discrete optimization methods for the large scale problem of optical flow.
Q11. What is the way to define the probability of assigning a given label to a?
After convergence of the messages, they can be used to define the probability of assigning a given label to a node, and the label with the maximum probability is chosen.
Q12. What is the way to penalize higher-order derivatives of the flow?
Another possibility is to penalize higher-order derivatives of the flow, as done in [241] for the second derivative, to favor piecewise affine flow fields.
Q13. How long does it take to process a pair of 640 480 pixels?
The related layered approach [233,234] achieving state-of-the-art results requires several hours to process a pair of 640 480 pixels, and even GPU-based implementation [244] can need up to an hour.