【CVPR 2020 メタサーベイ】3D From a Single Image and Shape-From-Xcvpaper. challenge
「CVPR 2020 網羅的サーベイ」により作成された 3D From a Single Image and Shape-From-X エリアのメタサーベイ資料です。
CVPR 2020 網羅的サーベイ: http://xpaperchallenge.org/cv/survey/cvpr2020_summaries/listall/
cvpaper.challengeはコンピュータビジョン分野の今を映し、トレンドを創り出す挑戦です。論文サマリ作成・アイディア考案・議論・実装・論文投稿に取り組み、凡ゆる知識を共有します。2020の目標は「トップ会議に30+本投稿」することです。
http://xpaperchallenge.org/cv/
SSII2021 [SS2] Deepfake Generation and Detection – An Overview (ディープフェイクの生成と検出)SSII
This document provides an overview of deepfake generation and detection. It begins with an introduction to the author and their background and research interests. The rest of the document is outlined as follows: definitions of deepfakes, various deepfake generation techniques including face synthesis, manipulation, reenactment and swapping, and an overview of deepfake detection methods including commonly used datasets, image-based and video-based detection approaches.
4. パッシブ3Dセンシング
• 写真測量
• Visual SLAM
• Multiview stereo
Gengshan Yang, Joshua Manela, Michael Happold, Deva Ramanan, Hierarchical
Deep Stereo Matching on High-resolution Images, CVPR 2019
Xiang Gao, Rui Wang, Nikolaus Demmel and Daniel Cremers, LDSO: Direct
Sparse Odometry with Loop Closure, IROS 2018
Jonathan Starck, Adrian Hilton, Surface Capture for Performance-Based
Animation, pp. 21-31, vol. 27, IEEE Computer Graphics 2007,
4
14. Motion blur
Projection pattern Captured image
Projector
Object motion
Camera
Multiple pattern projection
(Our technique)
How to decompose the multiplexed pattern?
No blur
14
17. Objective function: stereo + speed
3D search for all the pixels
Coarse to fine approach
camera projector
1. Coarse: constant velocity
2. Fine: arbitrary velocity
17
18. System overview
Off-the-shelf cameras and projectors:
Digital Light Processing (DLP) projectors
Pattern switching is faster than cameras fps (~10,000Hz)
Projector
Camera
DLP projector 18
19. Comparison on depth estimation results on a plane
Evaluation
0.000
0.001
0.002
0.003
0.004
0.005
0.006
0.007
0.008
0.009
0.010
stoped slow normal fast
RMSE [m]
Speed
1
3
6
Kinect
1
3
6
Previous method
STOP Slow Fast
The number of patterns projected within an exposure time
[ICCV2017]
Our method
19
24. • Robust Active 3D Measurement Method against Bokeh Using Projector-
Camera System with Coded Aperture [Horita, Kawasaki et al., IEICE Trans.D,
2013]
- Install the coded aperture on projector
- DfD based on deconvolution
- Coded aperture stabilize DfD process
Coded Aperture
Changyin Zhou, Shree Nayar
ICCP’09
符号化開口をプロジェクタに使用
24
25. 25システムの構成
Actual implementation
レンズ & 符号化開口 LED光源
• Lens: achromatic lens
- Focal length:150.0mm, Diameter:50.0mm
• Code aperture:
- Zhou et.al., “What are Good Apertures for Defocus Debluring?” ICCP2009
- Light path through 70%
• Projector: LED 20x20
• CCD camera: Pointgrey JHF12M-5MP
25
26. 26逆畳み込みによるデプス推定
1. Project pattern with CA1. Project pattern with CA
2. Deconvolve with each depth parameters2. Deconvolve with each depth parameters
3. Compare similarity between original & deconvolved img.3. Compare similarity between original & deconvolved img.
4. Select depth of maximum similarity4. Select depth of maximum similarity
250mm 290mm 350mm
0.9similarity 0.2 0.4
250mm
Reconstruction algorithm
Convolution process
26
31. Active One-Shot Scan for Wide Depth Range Using a Light Field Projector Based on
Coded Aperture, Hiroshi Kawasaki, Satoshi Ono, Yuki Horita, Yuki Shiba, Ryo Furukawa,
Shinsaku Hiura; ICCV, 2015
Slit aperture
Line pattern
DfDとステレオを同時に行う手法
高周波パターンによる高周波ライトフィールド構築
31
33. Wide range reconstruction results
[H. Kawasaki, H. Shiba, S. Hiura and R.Furukawa ICCV 2015]
Reconstructed shapes
Top view
3D shapes
Top view
Captured image
33
35. Reconstruction results
[H. Kawasaki, H. Shiba, S. Hiura and R.Furukawa ICCV 2015]
35
Captured
images
3D shapes
(low resolution)
3D shapes
(high resolution)
49. Bundle adjustment approach for a projector-
camera system is difficult!!
• Bundle adjustment
• Depends on “3D points
observable for multiple frames”
• Problem:
• Projected features are NOT real
3D points observable for multiple
frames projector
Object
camera
No information for inter-frame correspondences
49
50. Active bundle adjustment:
[Kawasaki and Furukawa, ECCV18workshop, 3DV2019]
• The shapes should be
consistent for different frames.
1C
2C
2P
1P
Inconsistencies between frames comes from calibration errors
minimize differences by optimize parameters
50
51. Active bundle adjustment:
[Kawasaki and Furukawa, 3DV2019]
Regarded as
a single point.Regarded as
a single point.
)( 1,1 jpπ
1,1 jp
)( 2,2 jpπ
2,2 jp
1, jip : 3D point (frame i)
()π : Line-of-sight
projection
51