Presentation is loading. Please wait.

Presentation is loading. Please wait.

Depth Estimation via Sampling

Similar presentations


Presentation on theme: "Depth Estimation via Sampling"β€” Presentation transcript:

1 Depth Estimation via Sampling
Eliad Tsfadia

2 The at most π‘˜-levels Let 𝐿 be 𝑛 lines in the plane.
Every intersection point between two line is called a vertex. A point 𝑝 ∈ π‘™βˆˆπΏ 𝑙 is of level π’Œ if there are exactly π‘˜ lines of 𝐿 strictly below it. The π’Œ-level is a curve along the lines of 𝐿 that passes through points in level π‘˜.

3 Questions (Upper Bounds)
What is the number of vertices in the 0-level ? The 0-level as at most π‘›βˆ’1 vertices (each line might contribute at most one segment to the 0-level) What is the number of vertices in the π‘˜-level ? This is a surprisingly hard question. What is the number of vertices of level ≀ π‘˜ ?

4 Theorem 1: For π‘˜>1, The number of vertices in 𝐴(𝐿) of level ≀ π‘˜ is 𝑂(π‘›π‘˜).
Proof: Let 𝑉 β‰€π‘˜ = 𝑉 β‰€π‘˜ (𝐿) be the set of all vertices of level ≀ π‘˜ in 𝐴(𝐿) Pick a random sample 𝑅 of 𝐿, by picking each line to be in the sample with probability 1 π‘˜ . Observe that 𝐸[|𝑅|] = 𝑛 π‘˜ For any vertex π‘£βˆˆπ΄(𝐿), Let 𝑋 𝑣 be an indicator variable which is 1 if 𝑣 is a vertex of 0-level in 𝐴(𝑅), and 0 otherwise. For π‘£βˆˆ 𝑉 β‰€π‘˜ (π‘Žπ‘ π‘ π‘’π‘šπ‘’ 𝑣 is a vertex in level j, for 0β‰€π‘—β‰€π‘˜) what is the probability that 𝑋 𝑣 = 1 ? Pr 𝑋 𝑣 =1 =( 1βˆ’ 1 π‘˜ ) 𝑗 ( 1 π‘˜ ) 2 β‰₯ ( 1βˆ’ 1 π‘˜ ) π‘˜ 1 π‘˜ 2 β‰₯ 𝑒 βˆ’2 π‘˜ π‘˜ π‘˜ 2 = 1 𝑒 2 π‘˜ 2 (note that 1βˆ’π‘₯β‰₯ 𝑒 βˆ’2π‘₯ for 0<π‘₯≀ 1 2 )

5 We saw that for π‘£βˆˆ 𝑉 β‰€π‘˜ , 𝐸 𝑋 𝑣 =Pr 𝑋 𝑣 =1 β‰₯ 1 𝑒 2 π‘˜ 2
On the other hand, the number of vertices on the 0-level of A(𝑅) is at most |𝑅|βˆ’1. As such: π‘£βˆˆ 𝑉 β‰€π‘˜ 𝑋 𝑣 ≀ 𝑣 ∈𝐴(𝐿) 𝑋 𝑣 ≀ 𝑅 βˆ’1 In expectation: E[ π‘£βˆˆ 𝑉 β‰€π‘˜ 𝑋 𝑣 ] ≀𝐸 𝑅 βˆ’1 = 𝑛 π‘˜ βˆ’1≀ 𝑛 π‘˜ On the other hand: E[ π‘£βˆˆ 𝑉 β‰€π‘˜ 𝑋 𝑣 ]= π‘£βˆˆ 𝑉 β‰€π‘˜ 𝐸 𝑋 𝑣 β‰₯ | 𝑉 β‰€π‘˜ | 𝑒 2 π‘˜ 2 Finally, we get: | 𝑉 β‰€π‘˜ | 𝑒 2 π‘˜ 2 ≀ 𝑛 π‘˜ . Namely, |𝑉 β‰€π‘˜ | ≀ 𝑒 2 π‘›π‘˜=𝑂(π‘›π‘˜) ∎

6 Extention of Theorem 1 - Clarkson Shor (89):
Suppose we have a set 𝑆 of 𝑛 objects (lines). The objects defines configurations (vertices). Each configuration is defined by at most 𝑑 objects (𝑑=2). Each configuration is in conflict with some of the objects (a vertex is in conflict with all the lines that below it) The weight/depth of a configuration is the number of objects that are in conflict with it (The level of a vertex). 𝑁 β‰€π‘˜ 𝑆 = The number of configurations in 𝑆 with weight β‰€π‘˜ (| 𝑉 β‰€π‘˜ |). 𝑁 β‰€π‘˜ 𝑛 = π‘šπ‘Žπ‘₯ 𝑆 =𝑛 𝑁 β‰€π‘˜ (𝑆) (the maximum of | 𝑉 β‰€π‘˜ (𝐿)| for every 𝐿 =𝑛). Then: 𝑁 β‰€π‘˜ 𝑛 =𝑂 π‘˜ 𝑑 𝑁 0 𝑛 π‘˜ ( 𝑁 β‰€π‘˜ =𝑂 π‘˜ 2 𝑁 0 𝑛 π‘˜ =𝑂 π‘˜ 2 βˆ™ 𝑛 π‘˜ =𝑂(π‘›π‘˜) ).

7 Proof (Reproduce of the proof of Theorem 1):
Let 𝑆 be a set of 𝑛 objects. Let 𝐢 β‰€π‘˜ = 𝐢 β‰€π‘˜ (𝑆) be the set of all configurations of weight ≀ k . Pick a random sample 𝑅 of 𝑆, by picking each line to be in the sample with probability 1 π‘˜ . Observe that 𝐸 𝑅 = 𝑛 π‘˜ . For any configuration 𝑐, Let 𝑋 𝑐 be an indicator variable which is 1 if 𝑐 is a configuration of weight 0 in 𝑅, and 0 otherwise. For π‘βˆˆ 𝐢 β‰€π‘˜ (assume 𝑐 is a configuration with weight 𝑗 in S, for 0β‰€π‘—β‰€π‘˜) what is the probability that 𝑋 𝑐 = 1 ? Pr 𝑋 𝑐 =1 =( 1βˆ’ 1 π‘˜ ) 𝑗 ( 1 π‘˜ ) 𝑑 β‰₯ ( 1βˆ’ 1 π‘˜ ) π‘˜ 1 π‘˜ 𝑑 β‰₯ 𝑒 βˆ’2 π‘˜ π‘˜ π‘˜ 𝑑 = 1 𝑒 2 π‘˜ 𝑑 (note that 1βˆ’π‘₯β‰₯ 𝑒 βˆ’2π‘₯ for 0<π‘₯≀ 1 2 )

8 We saw that 𝐸 𝑋 𝑐 =Pr 𝑋 𝑐 =1 β‰₯ 1 𝑒 2 π‘˜ 𝑑
As such: π‘βˆˆ 𝐢 β‰€π‘˜ 𝑋 𝑐 ≀ 𝑐 ∈ 𝐢 ≀𝑛 𝑋 𝑐 = 𝑁 0 𝑅 ≀ 𝑁 0 ( 𝑅 ) In expectation: E[ π‘βˆˆ 𝐢 β‰€π‘˜ 𝑋 𝑐 ] ≀𝐸 𝑁 0 (|𝑅|) On the other hand: E[ π‘βˆˆ 𝐢 β‰€π‘˜ 𝑋 𝑐 ]= π‘βˆˆ 𝐢 β‰€π‘˜ 𝐸 𝑋 𝑐 β‰₯ | 𝐢 β‰€π‘˜ | 𝑒 2 π‘˜ 𝑑 We get: | 𝐢 β‰€π‘˜ | 𝑒 2 π‘˜ 𝑑 ≀𝐸 𝑁 0 (|𝑅|) . Namely, 𝑁 β‰€π‘˜ 𝑆 =|𝐢 β‰€π‘˜ | = 𝑂( π‘˜ 𝑑 𝐸[ 𝑁 0 |𝑅| ]) Does 𝐸 𝑁 0 |𝑅| =𝑂 𝑁 0 𝑛 π‘˜ ?

9 Lemma: If 𝑁 0 𝑛 =𝑂(π‘π‘œπ‘™π‘¦ 𝑛 ) , Then 𝐸 𝑁 0 |R| =𝑂 N 0 n k .
If we assume the condition in this lemma, we get the final result: 𝑁 β‰€π‘˜ 𝑆 =𝑂 π‘˜ 𝑑 𝑁 0 𝑛 π‘˜ This is a bound for every 𝑆. Thus, 𝑁 β‰€π‘˜ 𝑛 =𝑂 π‘˜ 𝑑 N 0 n k ∎

10 Clarkson-Shor: 𝑁 β‰€π‘˜ 𝑛 =𝑂 π‘˜ 𝑑 𝑁 0 𝑛 π‘˜
Example: Let π‘ƒβŠ† ℝ 2 be a set of 𝑛 points above the π‘₯-axis. Let β„› be the set of rectangles that their bottom edge is lying on the π‘₯-axis and that contain a point of 𝑃 on each of their other three edges. What is the number of rectangle in β„› that contain at most π‘˜ points of 𝑃? Solution: Can we bound 𝑁 0 𝑛 ? (the number of empty rectangles) 𝑁 0 𝑛 ≀𝑛=𝑂(𝑛) (each point π‘βˆˆπ‘ƒ can have at most one such empty rectangle containing p on its upper edge) By Clarkson-Shor we get: 𝑁 β‰€π‘˜ 𝑛 =𝑂 π‘˜ 3 βˆ™ 𝑛 π‘˜ =𝑂( π‘˜ 2 𝑛) .

11 Another Example: Let 𝑃 be a set of 𝑛 points in ℝ 𝑑
Another Example: Let 𝑃 be a set of 𝑛 points in ℝ 𝑑 . A region is a halfspace with 𝑑 points on its boundary (the halfspace above the points). What is the number of different halfspaces containing at most π‘˜ points of 𝑃 ? Solution: Theorem without proof: the complexity of the convex hull of 𝑛 points in 𝑑 dimensions is bounded by 𝑂( 𝑛 𝑑 2 ). By the theorem we get: 𝑁 0 𝑛 =𝑂( 𝑛 𝑑 2 ). By Clarkson-Shor we get: 𝑁 β‰€π‘˜ 𝑛 =𝑂 π‘˜ 𝑑 𝑛 π‘˜ 𝑑 2 =𝑂 π‘˜ 𝑑 2 𝑛 𝑑

12 We saw that by Clarkson-Shor we can give a good upper bound for 𝑁 β‰€π‘˜ (𝑛) in many different problems, If we have a good bound for 𝑁 0 (𝑛). What about 𝑁 π‘˜ (𝑛)? If we stay at the first example (the π‘˜-level problem), can we give a good upper bound for the number of vertices in the π‘˜-level which is smaller than 𝑂(π‘›π‘˜)? We will see later a bound of 𝑂(𝑛 π‘˜ ) to the number of vertices in the π‘˜-level.

13 The Crossing Lemma Theorem 2 (Euler’s formula): For a connected planar graph 𝐺 with 𝑛 vertices, π‘š edges and 𝑓 faces we have: 𝑛+𝑓 =π‘š+2. If 𝐺 is not necessarily connected, we have: 𝑛+𝑓β‰₯π‘š+2. Lemma 3: If 𝐺 is a simple planar graph, then π‘šβ‰€3π‘›βˆ’6 and 𝑓≀2π‘›βˆ’4. Definition: For any simple graph 𝐺, denote as 𝑐(𝐺) as the minimal number of edge crossing in any drawing of 𝐺. Claim 4: For any simple graph 𝐺, 𝑐(𝐺) β‰₯π‘šβˆ’3𝑛+6

14 Claim 4: For any simple graph 𝐺, 𝑐(𝐺) β‰₯π‘šβˆ’3𝑛+6
Proof: Draw 𝐺 with 𝑐(𝐺) crossings. Let 𝐻 be the graph resulting from 𝐺 by removing one of the edges from each pair of crossing edges. We have π‘š 𝐻 β‰₯π‘šβˆ’π‘ 𝐺 . Note that 𝐻 is Planar. By Lemma 3 we get π‘š 𝐻 ≀3𝑛 𝐻 βˆ’6=3π‘›βˆ’6 We get: π‘šβˆ’π‘ 𝐺 ≀3π‘›βˆ’6 , Which is equivalent to 𝑐 𝐺 β‰₯π‘šβˆ’3𝑛+6. ∎

15 Lemma 5 (The Crossing Lemma): For a simple graph 𝐺 with π‘šβ‰₯6𝑛 we have 𝑐 𝐺 =Ξ©( π‘š 3 𝑛 2 ).
Proof: Draw 𝐺 with 𝑐(𝐺) crossings, and let π‘ˆ be a random subset of 𝑉(𝐺) selected by choosing each vertex with probability 𝑝 (We will choose 𝑝 later). Let 𝐻= 𝐺 π‘ˆ =(π‘ˆ, 𝐸 β€² ) be the induced subgraph of 𝐺 over π‘ˆ. 𝐸 β€² = 𝑒𝑣 π‘’π‘£βˆˆπΈ 𝐺 and 𝑒,π‘£βˆˆπ‘ˆ . What is the Probability that a vertex 𝑣 survived? Exactly 𝑝 What is the probability that an edge of 𝐺 survived? Exactly 𝑝 2 What is the probability of a crossing to survived? Exactly 𝑝 4

16 Let 𝑁 𝑣 , 𝑁 𝑒 and 𝑁 𝑐 denote the number of vertices, edges and crossing that survived, respectively. Then, 𝐸 𝑁 𝑣 =𝑛𝑝, 𝐸 𝑁 𝑒 =π‘š 𝑝 2 , 𝐸 𝑁 𝑐 =𝑐 𝐺 𝑝 4 By Claim 4 we have: 𝑁 𝑐 β‰₯𝑐 𝐻 β‰₯ 𝑁 𝑒 βˆ’3 𝑁 𝑣 +6 In particular: 𝐸 𝑁 𝑐 β‰₯𝐸 𝑁 𝑒 βˆ’3𝐸 𝑁 𝑣 +6 This implies that: 𝑐 𝐺 𝑝 4 β‰₯π‘š 𝑝 2 βˆ’3𝑛𝑝+6 We get: 𝑐 𝐺 β‰₯ π‘š 𝑝 2 βˆ’ 3𝑛 𝑝 𝑝 4 β‰₯ π‘š 𝑝 2 βˆ’ 3𝑛 𝑝 3 In particular, by setting 𝑝= 6𝑛 π‘š ≀1 we have that: 𝑐 𝐺 β‰₯ π‘š 6𝑛 π‘š 2 βˆ’ 3𝑛 6𝑛 π‘š 3 = π‘š 𝑛 2 βˆ’ π‘š 𝑛 2 = π‘š 𝑛 2 =Ξ© π‘š 3 𝑛 2 ∎

17 Duality A point 𝑝=(π‘Ž,𝑏) is transferred to the line 𝑝 βˆ— ={𝑦=π‘Žπ‘₯+𝑏} .
A line 𝑙={𝑦=𝑐π‘₯+𝑑} is transferred to the point 𝑙 βˆ— =(βˆ’π‘,𝑑) . We can extent this duality to dill with lines in the form {π‘₯=π‘Ž} in the projective plane with Homogeneous coordinates. For the sake of simplicity, assume there aren’t such lines. 𝑝= π‘Ž,𝑏 is on 𝑙= 𝑦=𝑐π‘₯+𝑑 ⟺ π‘Žπ‘+𝑑=π‘βŸΊβˆ’π‘Žπ‘+𝑏=π‘‘βŸΊ 𝒍 βˆ— = βˆ’π‘,𝑑 is on 𝑝 βˆ— ={𝑦=π‘Žπ‘₯+𝑏} 𝑝= π‘Ž,𝑏 is above 𝑙= 𝑦=𝑐π‘₯+𝑑 ⟺ π‘Žπ‘+𝑑<π‘βŸΊβˆ’π‘Žπ‘+𝑏>π‘‘βŸΊ 𝑝 βˆ— = 𝑦=π‘Žπ‘₯+𝑏 is above 𝒍 βˆ— = βˆ’π‘,𝑑

18 On the number of π’Œ-sets (Application of the crossing lemma)
Let 𝑃 be a set of 𝑛 points in general position. A pair of points 𝑝,π‘žβˆˆπ‘ƒ form a π‘˜-set if there are exactly π‘˜ points in the (closed) halfplane below the line passing through p and q. Consider the graph 𝐺=(𝑃,𝐸) that has an edge for every π‘˜-set. We want to bound the size of 𝐸 as a function of 𝑛 and π‘˜. Observe that via duality we have that the number of π‘˜-sets is exactly the number of vertices in the (π‘˜βˆ’2)βˆ’level in the dual arrangement 𝐴( 𝑃 βˆ— ). Thus, if we can bound the size of 𝐸, we also can bound the number of vertices in the π‘˜-level.

19 Lemma 6: Let π‘žπ‘ and 𝑠𝑝 be two π‘˜-set edges of 𝐺, with π‘ž and 𝑠 to the left of 𝑝. Then there exists a point π‘‘βˆˆπ‘ƒ to the right of 𝑝 such that 𝑝𝑑 is a π‘˜-set, and 𝑙𝑖𝑛𝑒(𝑝,𝑑) lies between 𝑙𝑖𝑛𝑒(π‘ž,𝑝) and 𝑙𝑖𝑛𝑒(𝑠,𝑝) Proof sketch: Below (and include) each of 𝑙𝑖𝑛𝑒(π‘ž,𝑝) and 𝑙𝑖𝑛𝑒(𝑠,𝑝) there are exactly π‘˜ points of 𝑃. If we increase by small πœ€ the slope of 𝑙𝑖𝑛𝑒(π‘ž,𝑝) (around 𝑝) we get that there are exactly π‘˜βˆ’1 points below this line (because it won’t include π‘ž) If we decrease by small πœ€ the slope of 𝑙𝑖𝑛𝑒(𝑠,𝑝) (around 𝑝) we get that there are exactly π‘˜ points below this line (because it will still include 𝑠). If we continue increase the slope of 𝑙𝑖𝑛𝑒(π‘ž,𝑝), we have to get from value π‘˜βˆ’1 to value π‘˜ somewhere before we get to 𝑠. These means that we got to a new point 𝑑 that has to be on the right of 𝑝 (why?) and below 𝑙𝑖𝑛𝑒(𝑝,𝑑) there are exactly π‘˜ points.

20 Lemma 7: Let π‘βˆˆπ‘ƒ and Let π‘žβˆˆπ‘ƒ be a point to the left of 𝑝 so that π‘žπ‘βˆˆπΈ .
Furthermore, assume that there are at least π‘˜βˆ’1 points of 𝑃 to the right of 𝑝. Then, there exists a point π‘‘βˆˆπ‘ƒ such that π‘π‘‘βˆˆπΈ and 𝑝𝑑 has larger slope than π‘žπ‘. Proof sketch: Below (and include) 𝑙𝑖𝑛𝑒(π‘ž,𝑝) there are exactly π‘˜ points of 𝑃. If we increase by small πœ€ the slope of 𝑙𝑖𝑛𝑒(π‘ž,𝑝) (through 𝑝) we get value π‘˜βˆ’1 If we increase the slope to infinity, we get value β‰₯ k. If we continue increasing the slope from π‘žπ‘ we have to get from value π‘˜βˆ’1 to π‘˜. These means that we got to a new point 𝑑 in the right of 𝑝 and below 𝑙𝑖𝑛𝑒(𝑝,𝑑) there are exactly π‘˜ points, and slope(𝑝𝑑) > slope(π‘žπ‘)

21 Lemma 8: The edges of 𝐺 can be decomposed into π‘˜βˆ’1 convex chains 𝐢 1 ,…, 𝐢 π‘˜βˆ’1 .
Proof (by iterative algorithm): The algorithm will have π‘˜βˆ’1 iterations. In each iteration, choose the left most point that still has an edge in 𝐸 that hasn’t been chosen yet. Suppose the point is π‘ž and the edge is 𝑒 = π‘žπ‘βˆˆπΈ. Rotate a line around 𝑝 (counterclockwise) till we encounter an edge 𝑒’ =π‘π‘ βˆˆπΈ where s is a point to the right of 𝑝. We can now walk from 𝑒 to 𝑒’ and continue walking in this way, forming a chain of edges in 𝐺. Claim: in the end of the algorithm, we get π‘˜βˆ’1 edge disjoint convex chains that contain all the edges in 𝐺.

22

23 Claim: In the end of the algorithm, we get π‘˜βˆ’1 edge disjoint convex chains that contain all the edges in 𝐺. Proof: The chains are convex since we rotate counterclockwise as we walk along a chain. By Lemma 6, no two chains can be β€œmerged” into using the same edge. By Lemma 6, no two chains can end in the same point. By Lemma 7, such a chain can end only in the last π‘˜βˆ’1 points of 𝑃. If there was an edge 𝑒=π‘π‘žβˆˆπΈ that hasn’t been chosen, we could create a convex chain from this edge, in contradiction to (2) and (3).

24 Lemma 8: The edges of 𝐺 can be decomposed into π‘˜βˆ’1 convex chains 𝐢 1 ,…, 𝐢 π‘˜βˆ’1 .
Similary, the edges of 𝐺 can be decomposed into π‘›βˆ’π‘˜+1 concave chains 𝐷 1 ,…, 𝐷 π‘›βˆ’π‘˜+1 . Proof of second part: Rotate the plane by 180Β°. Every π‘˜-set is now (π‘›βˆ’π‘˜+2)-set. By the first argumentation, the edges of 𝐺 can be decomposed into π‘›βˆ’π‘˜+1 convex chains, which are concave in the original orientation. ∎

25 Theorem 9: 𝐸 =𝑂(𝑛 π‘˜ ), which means that the number of π‘˜-sets defined by a set of n points in the plane is 𝑂(𝑛 π‘˜ ). Proof : The graph 𝐺 has 𝑃 =𝑛 vertices. Let π‘š=|𝐸 𝐺 | be the number of π‘˜-sets. By Lemma 8, the edges can be decomposed into convex chains 𝐢 1 ,…, 𝐢 π‘˜βˆ’1 and can be decompose also to concave chains 𝐷 1 ,…, 𝐷 π‘›βˆ’π‘˜+1 . Any crossing of two edges of 𝐺 is an intersection point of one convex chain of 𝐢 1 ,…, 𝐢 π‘˜βˆ’1 to one concave chain of 𝐷 1 ,…, 𝐷 π‘›βˆ’π‘˜+1 . Since a convex chain and a concave chain can have at most two intersections, we conclude that there are at most 2 π‘˜βˆ’1 π‘›βˆ’π‘˜+1 =𝑂(π‘›π‘˜) crossing in 𝐺. By the crossing lemma (Lemma 5) there are at least Ξ© π‘š 3 𝑛 2 crossings in 𝐺. Putting this two together, we conclude π‘š 3 𝑛 2 =𝑂 π‘›π‘˜ β‡’π‘š=𝑂(𝑛 π‘˜ ) ∎

26 On the number of incidences (Another application of the crossing lemma)
Let 𝑃 be a set of 𝑛 distinct points in the plane. Let 𝐿 be a set of π‘š distinct lines in the plane (we do not assume general position here). Let 𝐼(𝑃,𝐿) denote the number of point/line pairs (𝑝,𝑙) where π‘βˆˆπ‘ƒ , π‘™βˆˆπΏ and π‘βˆˆπ‘™ . Let 𝐼 𝑛,π‘š = π‘šπ‘Žπ‘₯ 𝑃 =𝑛 , 𝐿 =π‘š 𝐼(𝑃,𝐿) Clearly, 𝐼 𝑛,π‘š β‰€π‘›π‘š. Lemma 6: 𝐼 𝑛,π‘š =𝑂( 𝑛 π‘š 𝑛+π‘š)

27 Lemma 10: 𝐼 𝑛,π‘š =𝑂( 𝑛 π‘š 𝑛+π‘š) Proof: Let 𝑃 and 𝐿 be the set of 𝑛 points and π‘š lines realizing 𝐼(𝑛,π‘š). Note that each line contain at least one point, and 𝐼 𝑛,π‘š β‰₯max⁑(𝑛,π‘š). Let 𝐺 be a graph over the points of 𝑃. We connect two points if they lie consecutively on a common line of 𝐿. What is 𝑛(𝐺) and π‘š 𝐺 ? Clearly, 𝑛(𝐺)=𝑛 and π‘š 𝐺 =πΌβˆ’π‘š , where 𝐼=𝐼(𝑛,π‘š). Note that 𝑐(𝐺)≀ π‘š 2 (trivial bound) On the other hand, by lemma 5, if π‘š 𝐺 β‰₯6𝑛 𝐺 (i.e. πΌβˆ’π‘šβ‰₯6𝑛) We have 𝑐 𝐺 β‰₯ 𝑐 1 π‘š 𝐺 3 𝑛 𝐺 2 . Thus, 𝑐 1 (πΌβˆ’π‘š) 3 𝑛 2 = 𝑐 1 π‘š(𝐺) 3 𝑛(𝐺) 2 ≀𝑐(𝐺)≀ π‘š 2 Thus, 𝐼= 𝑐 1 βˆ’ 𝑛 π‘š π‘š=𝑂( 𝑛 π‘š π‘š)

28 We proved: π‘š 𝐺 β‰₯6𝑛 𝐺 ⇒𝐼 𝑛,π‘š =𝑂( 𝑛 2 3 π‘š 2 3 +π‘š).
If π‘š(𝐺)<6𝑛(𝐺) we get πΌβˆ’π‘šβ‰€6nβ‡’ 𝐼=𝑂 𝑛+π‘š Putting the two together, we have that 𝐼 𝑛,π‘š =𝑂( 𝑛 π‘š 𝑛+π‘š) ∎

29 Lemma 11: 𝐼 𝑛,π‘š =Ξ©( 𝑛 π‘š 𝑛+π‘š) Proof: It’s clear that 𝐼 𝑛,π‘š =Ξ© 𝑛+π‘š . For a positive integer k, let π‘˜ = 1,…,π‘˜ . Denote 𝑁 1 = 1 2 𝑛 π‘š βˆ’ 1 3 , 𝑁 2 = 𝑛 βˆ’ π‘š , 𝑀= 𝑛 π‘š For the sake of simplicity of exposition, assume 𝑁 1 , 𝑁 2 , 𝑁 3 are integers. Let 𝑃= 𝑁 1 X 2𝑀 and L= 𝑦=π‘Žπ‘₯+𝑏 π‘Žβˆˆ 𝑁 2 and π‘βˆˆ 𝑀 . Note that 𝑃 = 𝑁 1 βˆ™2𝑀=𝑛 and |𝐿|= 𝑁 2 βˆ™π‘€=π‘š. For any π‘₯∈ 𝑁 1 , and for any π‘Žβˆˆ 𝑁 2 and π‘βˆˆ 𝑀 we have that: 𝑦=π‘Žπ‘₯+𝑏≀ 𝑁 1 𝑁 2 +𝑀= 1 2 𝑀+𝑀≀2𝑀. Namely, any line of 𝐿 is incident to 𝑁 1 points of 𝑃. Namely, 𝐼 𝑃,𝐿 = 𝑁 1 𝐿 = 1 2 𝑛 π‘š βˆ’ 1 3 βˆ™π‘š= 1 2 𝑛 π‘š β‡’ 𝐼 𝑛,π‘š =Ξ©( 𝑛 π‘š ) ∎


Download ppt "Depth Estimation via Sampling"

Similar presentations


Ads by Google