Download presentation
Presentation is loading. Please wait.
Published byJocelyn Horn Modified over 9 years ago
1
Kourosh MESHGI Shin-ichi MAEDA Shigeyuki OBA Shin ISHII 18 MAR 2014 Integrated System Biology Lab (Ishii Lab) Graduate School of Informatics Kyoto University meshgi-k@sys.i.kyoto-u.ac.jp IEICE NC Tamagawa’14
2
KOUROSH MESHGI – ISHII LAB - DEC 2013 - SLIDE 2 MAIN APPLICATIONS Surveillance Public Entertainment Robotics Video Indexing Action Recog.
3
KOUROSH MESHGI – ISHII LAB - DEC 2013 - SLIDE 3 MAIN CHALLENGES Varying Scale Clutter Non-Rigid Illumination Abrupt Motion
4
[Mihaylova et al., 07] RGB Color + Texture + Motion + Edge, Two PF [Spinello & Arras,11] HOG on RGB + HOG on Depth, SVM Classification [Shotton et al, 11] Skeleton from Depth, Random Forrest [Choi et al, 11] Ensemble of Detectors (upper body, face, skin, shape from depth, motion from depth), RJ-MCMC [Song et al., 13] 2.5D Shape + Motion + HOG on Color and Depth, Occlusion Indicator, SVM
5
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 5 Frame: t Observation
6
Image Patch KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 6 Frame: t State w h (x,y)
7
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 7 Feature Set Color Shape Edge Texture
8
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 8 Frame: 1 Template f1f1 fjfj fnfn ……
9
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 9 Frame: 1 Particles Initialized Overlapped
10
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 10 Frame: t Motion Model → t + 1
11
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 11 Frame: t + 1 Feature Vectors f1f1 f2f2 fnfn X 1,t+1 X 2,t+1 X N,t+1 … …
12
KOUROSH MESHGI – ISHII LAB - MAR 2014- SLIDE 12 Frame: t Probability of Observation Each Feature Independence Assumption !
13
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 13 Frame: t + 1 Particles Brighter = More Probable
14
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 14 Frame: t + 1 Expectation
15
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 15 Frame: t + 1 New Model Model Update
16
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 16 Frame: t + 1 Proportional to Probability
17
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 17 Same Color Objects Background Clutter Illumination Change Shadows, Shades Use Depth!
18
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 18 Templates Corrupted! Handle Occlusion!
19
Particles Converge to Local Optima / Remains The Same Region KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 19 Advanced Motion Models (not always feasible) Restart Tracking (slow occlusion recovery) Expand Search Area!
20
do not address occlusion explicitly maintain a large set of hypotheses computationally expensive direct occlusion detection robust against partial & temp occ. persistent occ. hinder tracking GENERATIVE MODELS DISCRIMINATIVE MODELS Dynamic Occlusion: Pixels of other object close to camera Scene Occlusion: Still objects are closer to camera than the target object Apparent Occlusion: Due to shape change, silhouette motion, shadows, self-occ UPDATE MODEL FOR TARGET TYPE OF OCCLUSION IS IMPORTANT KEEP MEMORY VS. KEEP FOCUS ON THE TARGET Combine Them!
21
* The Search is not Directed * Neither of the Channels have Useful Information * Particles Should Scatter Away from Last Known Position KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 21 Occlusion!
22
Occlusion Flag (for each particle) Observation Model No-Occlusion Particles Same as Before Occlusion-Flagged Particles Uniform Distribution KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 22
23
Probability of Occlusion for the Next Box Modified Dynamics Model of Particle KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 23
24
Model Update Separately for each Feature KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 24
25
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 25 Occlusion!
26
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 26 Occlusion! GOTCHA!
27
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 27 Quick Occlusion Recovery Low CPE No Template Corruption No Attraction to other Object/ Background
28
COLOR (HOC) TEXTURE (LBP) EDGE (LOG) DEPTH (HOD) 3D SHAPE (PCL Σ ) KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 28
29
Princeton Tracking Dataset 5 Validation Video with Ground Truth 95 Evaluation Video KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 29
30
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 30 A. Edge PFB. Edge + Color PFC. Edge + Color + Depth PFD. Edge + Color + Depth + Texture PFE. Edge + Color + Depth + Texture + 3D Shape PFF. Occlusion Aware PF
31
(Yellow Dashed Line is Ground Truth)
32
PASCAL VOC toto Success Overlap Threshold 0 1 1 Area Under Curve KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 32
33
KOUROSH MESHGI – ISHII LAB - MAR 2014- SLIDE 33 Success Plot A D B E C F 1 1 Overlap Threshold Success Rate
34
Mean Central Point Error: Localization Success Mean Scale Adaption Error KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 34 EstimatedGround Truth
35
KOUROSH MESHGI – ISHII LAB - MAR 2014- SLIDE 35 Center Positioning Error A D B E C F 100 50 Frames CPE (pixels)
36
KOUROSH MESHGI – ISHII LAB - MAR 2014- SLIDE 36 Scale Adaptation Error KOUROSH MESHGI – ISHII LAB - MAR 2014- SLIDE 36 140 SAE (pixels) 50 Frames A D B E C F
37
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 37 Tracker AUCCPESAE A (edg) 15.72192.2741.62 B (edg+hoc) 29.8893.4746.71 C (edg+hoc+hod) 46.7434.6240.88 D (edg+hoc+hod+tex) 48.4930.1846.27 E (edg+hoc+hod+tex+shp) 58.0323.8429.62 F (all + occlusion handling) 63.5817.4625.07
38
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 38 More Resilient Features + Scale Adaptation Active Occlusion Handling Measure the Confidence of each Data Channel Adaptive Model Update
39
Q UESTIONS? Thank you for your time… Image Credit: http://www.engg.uaeu.ac.ae/
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.