KOUROSH MESHGI PROGRESS REPORT TOPIC To: Ishii Lab Members, Dr. Shin-ichi Maeda, Dr. Shigeuki Oba, And Prof. Shin Ishii 9 MAY 2014
KOUROSH MESHGI – ISHII LAB - DEC SLIDE 2 MAIN APPLICATIONS Surveillance Public Entertainment Robotics Video Indexing Action Recog.
KOUROSH MESHGI – ISHII LAB - DEC SLIDE 3 MAIN CHALLENGES Varying Scale Clutter Deformation Illumination Abrupt Motion
Goal: Define p(X t |Y 1,…,Y t ) given p(X 1 ) KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 4 States: Target Location and Scale Observations: Sensory Information
PARTICLE FILTER TR.
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 6 Frame: t RGB Domain
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 7 Frame: t Depth Domain
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 8 Frame: t Sensory Information
Observation KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 9 Frame: t State w h (x,y)
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 10 Feature Set Color Shape Edge Texture
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 11 Frame: 1 Template f1f1 fjfj fMfM ……
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 12 Frame: 1 Particles Initialized Overlapped
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 13 Frame: t Motion Model → t + 1
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 14 Frame: t + 1 Feature Vectors f1f1 f2f2 fMfM X 1,t+1 X 2,t+1 X N,t+1 … …
KOUROSH MESHGI – ISHII LAB - MAR SLIDE 15 Frame: t Probability of Observation Each Feature Independence Assumption !
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 16 Frame: t + 1 Particles Brighter = More Probable
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 17 Frame: t + 1 Feature Vectors
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 18 Frame: t + 1 New Model Model Update
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 19 Frame: t + 1 Proportional to Probability
PARTICLE FILTER TR.
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 21 Appearance ChangesModel DriftDeficient Feature SpaceUninformed SearchOptimized Feature SelectionApproximation of Target
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 22 Same Color Objects Background Clutter Illumination Change Shadows, Shades Use Depth!
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 23 Templates Corrupted! Handle Occlusion! (No Model Update During Them)
* Local Optima of Feature Space * Feature Noise * Feature Failures KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 24 Regularization Non-zero Values Normalization
Particles Converge to Local Optima / Remains The Same Region KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 25 Advanced Motion Models (not always feasible) Restart Tracking (slow occlusion recovery) Expand Search Area!
* The Search is not Directed * Neither of the Channels have Useful Information * Particles Should Scatter Away from Last Known Position KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 26 Occlusion!
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!
PTO partial occlusion SAO self- or articulation occlusion TFO temporal full occlusion - shorter than 3 frames PFO persistent full occlusion CPO complex partial occlusion - including “split and merge” and permanent changes in a key attribute of a part of target CFO complex full occlusion KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 28
[Zhao & Nevatia, 04] Occlusion Indicator: Ratio of FG/BKG [Wu & Nevatia, 07] Handle Occlusion using Appearance Model [de Villiers et al., 12] Switch Tracker in the case of Occlusion [Song & Xiao, 13] Occlusion Indicator: New Peak in HOD or Reduction of the Size of Main Peak Many other papers handle occlusions as the by- product of their novel trackers
OCCLUSION AWARE PFT
Motion Model Resampling Target Estimation Calculate Likelihood KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 31 Initialization Model Update Observation Occlusion Flag? Constant Likelihood Occlusion Estimation Occlusion Threshold > ? YES NO
Occlusion Flag (for each particle) Observation Model No-Occlusion Particles Same as Before Occlusion-Flagged Particles Uniform Distribution KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 32
Position Estimation of the Target Occlusion State for the Next Box KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE a x x
Model Update (Separately for each Feature) Modified Dynamics Model of Particle KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 34
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 35 Occlusion!
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 36 Occlusion! GOTCHA!
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 37 Quick Occlusion Recovery Low CPE No Template Corruption No Attraction to other Object/ Background
COLOR (HOC) TEXTURE (LBP) EDGE (LOG) 2D PROJ. (BETA) 3D SHAPE (PCL Σ ) DEPTH (HOD) GRADIENT (HOG) KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 38
& DISCUSSION
Princeton Tracking Dataset 5 Validation Video with Ground Truth 95 Evaluation Video KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 40
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 41 OAPFT (Proposed, with different feature sets) State-of-the-art RGBD tracker OI+SVM (SVM tracker with Occlusion Indicator) Traditional Particle Filter tracker ACPF (Adaptive Color Particle Filter) State-of-the-art RGB tracker, Successful for Occlusion Handling STRUCK (Structured Output SVM Tracker)
PASCAL VOC: Overall Performance toto Success Overlap Threshold Area Under Curve KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 42
KOUROSH MESHGI – ISHII LAB - MAR SLIDE Success Plot Overlap Threshold Success Rate 1 1
Mean Central Point Error: Localization Success Mean Scale Adaption Error KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 44 EstimatedGround Truth
Center Positioning Error Frames CPE (pixels)
Scale Adaptation Error Frames SAE (pixels)
FP happens when a tracker doesn’t realize that the target is occluded. MI happens when the target is visible but the tracker fails to track it as if the target is still in an occlusion state MT the estimated bounding box has nothing in common with ground truth box FPS execution time in frames per second KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 47
KOUROSH MESHGI – ISHII LAB – MAR 2014 – SLIDE 48 Tracker AUCCPESAEMIFPMTFPS BCDEGST (proposed) ACPF (Nummiaro ‘03) STRUCK (Hare ‘11) OI+SVM (Song ‘13)
KOUROSH MESHGI – ISHII LAB – MAY 2014 – SLIDE 49 More Resilient Features + Scale Adaptation Active Occlusion Handling Measure the Confidence of each Data Channel Adaptive Model Update
Q UESTIONS? Thank you for your time…