Download - Sparselet Models for Efficient Multiclass Object Detection

Transcript
Page 1: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

Sparselet Models for Efficient Multiclass Object Detection

Present by Guilin Liu

Page 2: Sparselet Models for Efficient Multiclass Object Detection

Key Idea

Use sparse coding of part filters to represent each filter as a sparse linear combination of shared dictionary elements.

Reconstruction of original part filter responses via sparse matrix-vector product

GPU implementation

masc.cs.gmu.edu

Page 3: Sparselet Models for Efficient Multiclass Object Detection

Problem/motivation

Individual model become redundant as the number of categories grow------Sparse Coding

Learn basis parts so reconstructing the response of a target model is efficient

masc.cs.gmu.edu

Page 4: Sparselet Models for Efficient Multiclass Object Detection

Overview

masc.cs.gmu.edu

System pipeline

Page 5: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

Overview

Page 6: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

1. Sparse reconstruction

Find a generic dictionary approximate the part filters pooled from a set of training models, subject to a sparsity constraint

Page 7: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

1. Sparse reconstruction

Solve the optimization problem busing the Orthogonal Matching Pursuit algorithm(OMP)Two steps:a.Fixed D, optimize αb.Fixex α, optimize D

Page 8: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

2. Precomputation & efficient reconstruction

Page 9: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

2. Precomputation & efficient reconstruction

1. Precompute convolutions for all sparselets2. Approximate t convolution response by linear

combination of the activation vectors from step 1.

Page 10: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

3. Implementation(CPU, GPU)

The independence and parallelizablity of:Convolution, HOG computation and distance transforms

1. CPU implementation: CPU cach miss limited the overall speedup

2. GPU implementation: a. Compute image pyramids and HOG featuresb. Compute filter responses to root, part or part basis

filter

Page 11: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

4. Experiments

1. Reconstruction error

Page 12: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

4. Experiments

2. held-out evaluation

Page 13: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu

4. Experiments

3. Average precision

Page 14: Sparselet Models for Efficient Multiclass Object Detection

masc.cs.gmu.edu