PittPatt Face Detection

1
PittPatt Face Detection Tracking CLEAR 2007

• Michael C. Nechyba
• Louis Brandy
• Henry Schneiderman
• Pittsburgh Pattern Recognition

2
System Overview

• Stage 1 Frame-based face detection
• Selective visual attention (new)
• PittPatt face finder (improved)
• Stage 2 Motion-based tracking
• Causal, second-order motion model (improved)
• Globally optimal data association (new)
• Stage 3 Track-filtering
• Merging of spatially consistent tracks
• Elimination of low-confidence tracks
• Adjustment of bounding boxes to conform to
  annotations
• No site or domain-specific prior knowledge
• Face finder was not custom-trained with
  development data
• All video was processed with identical
  configuration / parameters

http//demo.pittpatt.com
3
Emphasis on Speed Performance

• From 2005 to present, 500 speed up
• Improved hardware performance gt 1.5
• Individual processors
• Compilers
• Object detector speed-ups gt 10
• Algorithmic re-design for speed
• Code-level optimizations
• Parallel implementations for multi-core world 4
• Selective visual attention gt 6

2005
2007
4
Selective Visual Attention
selective visual attention
face tracking results
percent of image processed by face detector
(moving avg.)
clip avg. 16.4
pixels
time
5
Speed Performance

• Test platform
• Processor Dual 3GHz Intel Xeon 5160
• Memory 4GB RAM
• Hard drive 500GB SATA II (3.0 GB/s)
• Better than real-time performance across all data
  sets

6
2005 Present Speed Performance Improvement
per clip range of real-time processing factors
real-time
VACE/CLEAR Evaluations
7
Accuracy Performance
VACE-2 excludes overhead clips

• Dominant failure modes
• Small, poor-quality faces
• Poses outside range of detector (90
  out-of-plane rotation, 45 frontal rotation)
• Inter-site performance variations due to small
  face sizes, poses, video quality...

8
Inter-Site Accuracy Variations
AIT (90/11)
IBM (64/5)
ITC (79/12)
UKA (85/18)
UKA (87/7)
good quality video
very small faces
blurred video quality small faces
highly compressed small faces
highly interlaced some small faces
diverse scenes
9
Algorithmic Accuracy Differences
CLEAR 2006 vs. CLEAR 2007 (2006 test data)
full frame processing vs. selective visual
attention
83.3 _at_ 8.7 FAP vs. 79.0 _at_ 4.8 FAP
80.8 _at_ 9.7 FAP vs. 80.1 _at_ 9.1 FAP
AIT
AIT-06
UPC
UPC-06
UKA
detection percentage
detection percentage
VACE
VACE
ITC
IBM-06
IBM
false alarm percentage
false alarm percentage
10
Accuracy of Accuracy
examples scored as false alarms (48 of all
false alarms on CHIL 2007 test data)
visually ambiguous examples scored as misses