Life-Long Place Recognition by Shared Representative...

16
Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1 , Xue Yang 1 , Yiming Deng 2 , Mark Rentschler 3 , Dejun Yang 1 , and Hao Zhang 1 1. Division of Computer Science, Colorado School of Mines 2. Department of Electrical Engineering, University of Colorado Denver 3. Department of Mechanical Engineering, University of Colorado Boulder

Transcript of Life-Long Place Recognition by Shared Representative...

Page 1: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Life-Long Place Recognition by Shared Representative

Appearance Learning

Fei Han1, Xue Yang1, Yiming Deng2,

Mark Rentschler3, Dejun Yang1, and Hao Zhang1

1. Division of Computer Science, Colorado School of Mines

2. Department of Electrical Engineering, University of Colorado Denver

3. Department of Mechanical Engineering, University of Colorado Boulder

Page 2: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Introduction

This work addresses long-term place recognition with strong appearance variations

2

Spring Summer

Autumn Winter

Morning

Afternoon

October

December

Page 3: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Introduction

The Problem

3

• Which feature is more representative while insensitive to appearance changes?

• How to fuse these heterogeneous feature modalities?

Page 4: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Contributions

Shared Representative Appearance Learning (SRAL) model for long-term place recognition

Capabilities Autonomous discriminative feature modality

learning and fusion that is robust to strong appearance variations

Sparse optimization formulation regularized by structured sparsity-inducing norms

Efficient solver with theoretical convergence guarantee

4

Page 5: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

ApproachScene Representation

5

• A set of scenes is represented by

• Scene scenarios are represented as

𝒙𝒊 =𝒙𝒊𝟏

⋮𝒙𝒊𝒎∈ ℝ𝑑

Feature modality 1

𝒀 =

𝒚𝟏⋮𝒚𝒏∈ ℝ𝑛×𝑐

𝑿 = 𝒙𝟏 ⋯ 𝒙𝒊 ⋯ 𝒙𝒏 ∈ ℝ𝑑×𝑛

⋮Feature modality m

Page 6: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

ApproachShared Representative Appearance Learning

6

ℓ2,1-norm: enforces sparsity of each row of W

ℓ𝑀-norm: enforces sparsity of each feature modality

Formulation:

Page 7: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Approach Optimization Algorithm

Theorem:

The Algorithm

converges to the

global optimal solution

of the formulated

optimization problem

7

Page 8: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

ApproachPlace Recognition

• Optimal weight 𝜔𝑖 for each feature modality 𝑖, 𝑖 = 1, … ,𝑚

𝜔𝑖 =∥ 𝑾𝒊∗ ∥𝐹

• Feature fusion for image matching

𝑠(𝑞, 𝑝) =

𝑖=1

𝑚

𝜔𝑖𝑠𝑖 (𝑞, 𝑝)

8

Page 9: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Experiments

• Three public datasets Different time of the day:

St Lucia Dataset Different months: CMU-

VL Dataset Different seasons:

Nordland Dataset

• Setups Hardware: Workstation

with i7 3.4GHz, 16GB Platform: Ubuntu 14.04 +

Matlab 2015 Feature modalities: Color,

GIST, HOG, and LBP features

9

Page 10: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

ExperimentsSt Lucia Dataset (Various Times of the Day)

10

Qualitative Evaluation

Page 11: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

ExperimentsSt Lucia Dataset (Various Times of the Day)

11

Quantitative Evaluation

Page 12: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Experiments CMU-VL Dataset (Different Months)

12

Qualitative Evaluation

Page 13: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

ExperimentsCMU-VL Dataset (Different Months)

13

Quantitative Evaluation

Page 14: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

ExperimentsNordland Dataset (Different Seasons)

14

Qualitative Evaluation

Page 15: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Experiments Nordland Dataset (Different Seasons)

15

Quantitative Evaluation

Page 16: Life-Long Place Recognition by Shared Representative ...fhan/publication/slides/rssw16_sral...Life-Long Place Recognition by Shared Representative Appearance Learning Fei Han 1, Xue

Summary

We introduced a long-term place recognition method, named Shared Representative Appearance Learning(SRAL), which

• Can autonomously learn and integrate shared discriminative features that are robust to strong appearance variations

• Formulates the task as a convex optimization problem regularized by structured sparsity

• Can be efficiently solved with theoretical convergence guarantee

16

Life-Long Place Recognition byShared Representative Appearance Learning