Tracking-Learning-Detection(TLD) Zdenek Kalal, Krystian Mikolajczyk, - PowerPoint PPT Presentation

Tracking-Learning-Detection(TLD) Zdenek Kalal, Krystian Mikolajczyk, Jiri Matas PAMI 2010 Presented by: Lyne P. Tchapmi Stanford University/CS231B

Overview • Problem Definition • Previous Works • Hinterstossier et al. • Babenko et al. • Contribution • TLD Framework • Tracker • Detector • Learning • Performance Analysis

Problem Definition Object Bounding Box Location TRACKING Video Frame

Overview • Problem Definition • Previous Works • Hinterstossier et al. • Babenko et al. • Contribution • TLD Framework • Tracker • Detector • Learning • Performance Analysis

Hinterstossier et al. CVPR 2009 Sparse Bayesian Learning for Efficient Visual Tracking RVM: Relevance Vector Machine (modified SVM)

Grabner et al. ECCV 2008

Overview • Problem Definition • Previous Works • Hinterstossier et al. • Babenko et al. • Contribution • TLD Framework • Tracker • Detector • Learning • Performance Analysis

TLD Contribution • New Tracking Framework (Tracking-Learning-Detection TLD) • P-N Learning • Handles unknown objects • Long-term tracking • Detector resets tracker (avoids drift) • Detector improves over time

Overview • Problem Definition • Previous Works • Hinterstossier et al. • Babenko et al. • Contribution • TLD Framework • Tracker • Detector • Learning • Performance Analysis

Implementation: TLD in detail

Implementation: Tracker • Pyramidal Lucas-Kanade Tracker (KLT)

Implementation: TLD in details

Implementation: Object detector • Scanning-window grid + Cascaded classifier 2-bit binary patterns Randomized Fern Forest

Implementation: TLD in details

Implementation: TLD in details

Implementation: Memory 𝑞 + : 𝑝𝑐𝑘𝑓𝑑𝑢 𝑞 − : 𝑐𝑏𝑑𝑙𝑕𝑠𝑝𝑣𝑜𝑒

Implementation: P-N Learning

P-N-LEARNING

Implementation: P-Expert I’ve seen this before…. …add to model!

Implementation: P-Expert Update Core Initial Model Add points on valid trajectory(solid)

Implementation: N-Expert If it isn’t positive, it must be negative… …add to model!

Implementation: TLD in details

Overview • Problem Definition • Previous Works • Hinterstossier et al. • Babenko et al. • Contribution • TLD Framework • Tracker • Detector • Learning • Performance Analysis

Performance Metrics • Precision 𝑄 = #𝑢𝑠𝑣𝑓 𝑞𝑝𝑡 #𝑠𝑓𝑡𝑞𝑝𝑜𝑡𝑓𝑡 • Recall #𝑢𝑠𝑣𝑓 𝑞𝑝𝑡 𝑆 = #𝑝𝑑𝑑𝑣𝑠𝑠𝑓𝑜𝑑𝑓𝑡 𝑢𝑝 𝑚𝑏𝑐𝑓𝑚 • F-measure 𝐺 = 2𝑄𝑆 𝑄 + 𝑆

Performance: CoGD Dataset • TLD achieves maximal possible score in all sequences

Performance: Prost Dataset • TLD scores best in 9/10 sequences, outperforms 2 nd best by 12%

Performance: TLD Dataset

Performance: TLD Dataset • TLD scores best on average 81%, vs 22% for 2 nd best (F-measure)

Limitations • Articulated Objects (pedestrians) • Full out of plane rotations • Tracker gets lost • Detector sees an appearance never seen in model • Tracker is fixed • Makes the same errors

QUESTIONS?