Bullet Time 1,a) 1 Bullet Time Bullet Time Generation Technique and Eveluation on High-Resolution Bullet-Time Camera Work Ryuuki Sakamoto 1,a) Ding Chen 1 Abstract: The multi-camera environment have been used in movie studio when they intent to apply the Bullet Time camera work to a scene. The camera work is realized with lipping through rames at same moment taken by multi cameras surrounding an object at even distances. For making outcome rames o the camera work, the Homography transormation is adapted or rectiying inaccurate camera poses. Thereore the Homography transormation, however, makes some blank spaces during distorting the rame, the scale up transormation should be applied ater that. The scaling up, however, makes the quality o the outcame down. In this paper, we proposed a method to calculate Homography matrices or keeping the quality and naturality o outcome rames o the camera work. For measuaring the eectiveness o the method, we also describe the result o a user evaluation. 1. 3 1 Bullet Time 10 Bullet Time [1] 1 Yahoo Japan Corp. Mid9-7-1 Akasaka, Minato-ku a) ryusakam@yahoo-corp.jp c 2014 Inormation Processing Society o Japan 1
8 12 30ps XGA VGA 1 3 FullHD TV PC Fig. 1 Conversion by eq.3 Bullet Time 3 Web [3], [4] Bullet Time H k k = 1,...,N Bullet Time k A k [2] R k T k A k Bullet Time k 0 u0 k A k = 0 k v0 k (1) 2. Bullet Time 0 0 1 xyz xz y = 0 y k u0 k v0 k N C k = R 1 k Bullet Time T k G e z e z y 3 G x e x e z e x z e z G R k G xz G k C k k = 1,...,N e x / e x C k C k R k = e y / e y (2) G e z / e z H k H k = A k R kr 1 k 3 A 1 k (3) C k G G 1 G c 2014 Inormation Processing Society o Japan 2
H k = A kr kr 1 k A 1 k (5) 4. 2 Fig. 2 Conversion by eq.3 2 v αk (α = 0,1,2,3) H k v βk (β = 0,1,2,3) [ ] [ ] v βk λ = H v αk k (6) G 1 1 g k G 2 v αk v βk 2 v βk S k S0 S0 / S k S k k (u0 k,v0 k ) S0 H k 3. S k = 1 3 3 (v(β mod 4)k 8 v αk) (7) α=0β=0 (v (β+1 mod 4)k v αk ) 1 k E C k G N k argmax E( k,u0 k,v0 k ) = S0 (8) C k G G S k G C k G G C k E k (u0 k,v0 k ) [1] Bullet Time (u0 k,v0 k ) C k R k 1: (u0 k,v0 k ) (u0 k,v0 k ) G G g k k N E k=1 g k/n (u0 k,v0 k ) A k 2: (u0 k,v0 k ) H k 1 (u0 k,v0 k ) g k k 0 u0 k k A k = 0 k v0 k (4) N g k 0 0 1 k=1 5. c 2014 Inormation Processing Society o Japan 3
g k 1 1 E 3: k 1 A B k 2 C D k C k G = average (9) z average N k average = N z average = k=1 N k=1 C k G N 1 Eye Bullet Time Vision[12] Eye Vision G 4: g k 2 Bullet Time C k G C k C k [1] k Eye Vision α k k α = C k G k = α(c k G) (10) 2 1 (u0 k,v0 k ) α 2 3 A k 3 k k E 3 E (1) G (2) 1 2 (u0 k,v0 k ) (3) 3 4 k (5) H k 1 2 3 4 3 4 B 3 k 1 4 4 2 A B C D Table 1 Combination o each technique 3 4 6. [5], [6], [7], [8], [9], [10], [11] Bullet Time 3 TV 7. 8 3 1 (4) (7) k 1 3 A 2 Bullet Time c 2014 Inormation Processing Society o Japan 4
3 Bullet Time Fig. 3 Bullet Time camera work on each scene 2 E Table 2 E values o each technique. A B C D 0.33 0.60 0.72 0.78 0.26 2 E 0.33 0.60 4 α A B α B α C A 1 2 (u0 k,v0 k ) 4 2 4 A B α Fig. 4 α on combination A and B E 0.72 5 (u0 k,v0 k ) A C (u0 k,v0 k ) C 5 k (u0 k,v0 k ) D E 0.78 5 A C (u0 k,v0 k ) Fig. 5 (u0 k,v0 k ) on combination A and C 8. Bullet Time E k (u0 k,v0 k ) 8.1 Bullet Time Point Grey Research Flea2 Web 8 12 4 90 150 c 2014 Inormation Processing Society o Japan 5
30 90 A D http Web HTML5 JavaScript JavaScript 80 Bullet Time 6 4 A Bullet Time 7 Web Fig. 7 Screenshot o an evaluation page 3 5 Table 3 Result o telepresence. Web A B C D Bullet Time n 173 142 113 142 4.82 5.23 5.03 5.30 7 1 2.04 1.30 1.89 1.35 T 7.55 12.93 8.00 11.52 5.08 1.80 T 19.34 Bullet Time Bullet Time Bullet Time 8.2 1 Bullet Time 8.2.1 1 Bullet Time 7 5 Web 4 6 4 A c 2014 Inormation Processing Society o Japan 6
6 Bullet Time Fig. 6 Bullet Time camera work on each scene 8.2.2 1 7 3 Bullet Time A D 1% Bullet Time 4 t Bullet Time 1% 8.3 2 4 F(3,566)=4.23, p<.01 Tukey 5% A D 8.3.1 A B A D 5 Web 1 D c 2014 Inormation Processing Society o Japan 7
4 4 Table 4 Results o each technique. A B C D 4 E 0.74 0.77 0.82 0.75 1 4.12 4.17 3.79 4.13 2.05 2.03 2.59 2.72 1 A 2 4.14 4.31 3.72 3.71 Bullet Time 1.64 2.57 1.84 2.21 B C D A Bullet Time 1 1 Bullet Time 4 1 [1] : 2 Vol. 106, No. 429, pp. 43 48 (2006). 1 [2] Nakanishi, H., Murakami, Y. and Kato, K.: Movable cameras enhance social telepresence in media spaces, Proceedings o the SIGCHI, ACM, pp. 433 442 (2009). 1 [3] Hartley, R. and Zisserman, A.: Multiple view geometry in computer vision, Cambridge Univ Press (2000). [4] Szeliski, R.: Computer vision: algorithms and applications, Springer (2011). 2 [5] Kanade, T., Rander, P. and Narayanan, P.: Virtualized reality: Constructing virtual worlds rom real scenes, 8.3.2 MultiMedia, Vol. 4, No. 1, pp. 34 47 (1997). [6] Kitahara, I. and Ohta, Y.: Scalable 3D representation D or 3D video in a large-scale space, Presence: Teleoperators and Virtual Environments, Vol. 13, No. 2, pp. E 0.45 4 E 164 177 (2004). [7] Inamoto, N. and Saito, H.: Intermediate view generation 1 o soccer scene rom multiple videos, Proc. on 16th International Conerence on Pattern Recognition, Vol. 2, IEEE, pp. 713 716 (2002). (F(3,135)=0.48, n.s.) 2 [8] Hilton, A., Guillemaut, J., Kilner, J., Grau, O. and Thomas, G.: 3d-tv production rom conventional cameras or sports broadcast, Broadcasting, IEEE Transac- (F(3,135)=1.58, n.s.) tions on, Vol. 57, No. 2, pp. 462 476 (2011). [9] Hashimoto, T., Uematsu, Y. and Saito, H.: Generation o see-through baseball movie rom multi-camera views, IEEE International Workshop on Multimedia Signal Processing, IEEE, pp. 432 437 (2010). Bullet Time [10] Kimura, K. and Saito, H.: Video synthesis at tennis player viewpoint rom multiple view videos, Proceedings. VR 2005., IEEE, pp. 281 282 (2005). 9. [11] Tomiyama, K., Miyagawa, I. and Iwadate, Y.: Prototyping o HD Multi-Viewpoint Image Generating Bullet Time System-Live broadcasting use at gymnastics competition (60 th National Championships)-, IEIC Technical Report, Vol. 106, No. 429, pp. 43 48 (2006). 4 [12] Kanade, T. et al.: EyeVision, Web, http://www. pvi-inc. com/eyevision. 1 c 2014 Inormation Processing Society o Japan 8