Download presentation
Presentation is loading. Please wait.
Published byJulian Willis Modified over 9 years ago
1
HIGH PERFORMANCE OBJECT DETECTION BY COLLABORATIVE LEARNING OF JOINT RANKING OF GRANULES FEATURES Chang Huang and Ram Nevatia University of Southern California, Institute for Robotics and Intelligent Systems
2
Outline Introduction Granules JRoG features Incremental Feature Selection Method Simulated Annealing Collaborative Learning Dynamic Search for Bayesian Combination Experiments Conclusion
3
Introduction Detect pedestrians with part occluded people Speed up and Accuracy up Collaborate learning of Simulated Annealing and increment selection model Dynamic search to improve Bayesian combination
4
Granules
5
JRoG features (Joint Ranking of Granules)
6
JRoG example
7
Distance Definition
8
Neighbor
9
Incremental Feature Selection method (Z is normalization factor) N is number of training samples M is number of features Time complexity is O(M ln N), better than O(MN) in conventional AdaBoost
10
Simulated Annealing Heuristically set N=1000 x dim(g 0 ), r = 0.01 1/n Θ 1 =1 Θ 2 =8 So each granule can be changed 1000 times and SA ends at temperature 0.01T 0 Selection of initial temperature (T 0 is critical)
11
Flow Chart
12
Collaborative Learning
13
Joint Likelihood F: full body H: head and shoulder T: torso L: legs Z: detection responses S: state of multiple humans Wu and Nevatia[19] uses Bayesian combination to deal with partial occlusions in crowded scenes Wu and Nevatia’s search
14
Dynamic search
15
Experiment1 Collaborate learning CL: Jump/keep ratio = 1.0, 0.2, 0.25 Initial temp.= 0.03, JRoG # bit = 3 SL: without SA process Evaluate Score: EER (Equal Error Rate) FPR (False Positive Rate)
16
Experiment1
17
Experiment2 INRIA dataset Training: 2478 positive, 1218 negative samples from dataset 24780 positive by rotating, scaling above Testing: 1128 positive, 453 negative samples from dataset
18
Experiment3 ETHZ Dataset Four 640x480 videos (one for training, one for testing) 23000 negatives from internet More than 20000 pedestrians labeled Outperform others in all three videos
19
Experiments
22
Computational Cost Xeon 3GHz Takes 70ms to scan 640x480 ETHZ images at 16 scales from 1.0 to 0.125 Training of 16-layer cascade costs 2 days
23
Conclusion A novel collaborative learning method Dynamic Search method for Bayesian combination Improves efficiency and accuracy Extensive to other objects like cars and faces
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.