Convolutional Pose Machines @conta_
緒方 貴紀 (twitter: @conta_)
CTO@ABEJA, Inc.
Computer Visionとか、Machine Learningを使った
プロダクト開発をやっています。
Self Introduction
Related Work
Pictorial structures Hierarchical models
Sequential prediction Convolutional architectures
[A. Toshev and C. Szegedy, CVPR’2013]
[Tian et al., ICCV’2011][Mykhaylo et al., CVPR’2009]
[Ramakrishna et al., 2014]
Related Work
Pictorial structures Hierarchical models
Sequential prediction Convolutional architectures
[A. Toshev and C. Szegedy, CVPR’2013]
[Tian et al., ICCV’2011][Mykhaylo et al., CVPR’2009]
[Ramakrishna et al., 2014]
Pose Machines
[Ramakrishna et al., 2014]
パッチから特徴量を抽出し、各Parts or NotのClassifierを用いて、Confidence Mapsを作りたい
Part Contextは非常に有効な特徴
Pose Machines
[Ramakrishna et al., 2014]
前段階での推定結果を用いて、各Partsの関係性を
事前情報無しにを活用出来ないか?
Pose Machines
Stage I Confidence
Head Neck L-Shoulder L-Elbow L-Wrist
g2g1 g3
Context Features
Context Features
Stage I Confidence Maps
Stage II Confidence Maps
Stage III Confidence Maps
Image Features
[Ramakrishna et al., 2014]
Pose Machines
Stage II Confidence
g2g1 g3
Context Features
Context Features
Stage I Confidence Maps
Stage II Confidence Maps
Stage III Confidence Maps
Image Features
Head Neck L-Shoulder L-Elbow L-Wrist
[Ramakrishna et al., 2014]
Pose Machines
Stage III ConfidenceHead Neck L-Shoulder L-Elbow L-Wrist
g2g1 g3
Context Features
Context Features
Stage I Confidence Maps
Stage II Confidence Maps
Stage III Confidence Maps
Image Features
Head Neck L-Shoulder L-Elbow L-Wrist
[Ramakrishna et al., 2014]
Pose Machines
Stage III ConfidenceHead Neck L-Shoulder L-Elbow L-Wrist
g2g1 g3
Context Features
Context Features
Stage I Confidence Maps
Stage II Confidence Maps
Stage III Confidence Maps
Image Features
Head Neck L-Shoulder L-Elbow L-Wrist
[Ramakrishna et al., 2014]
各StageのConfidence Mapsと教師データとのEuclidean Distance Loss
教師データ: 各PartsのGround truth locationからGaussian Peakを計算したもの
Loss Function
3つのDatasetsで実験
- MPII Human Pose Dataset
- Leeds Sports Pose (LSP) Datase
- FLIC Dataset
Experiments
1. (色々頑張って実験した結果)いい感じの連続構成CNNによって、暗黙的な空間モデルの学習ができた
2. Graphical Modelによる推論無しに、階層構造のPredictionができた
Conclusion
We are hiring!
→ https://www.wantedly.com/companies/abeja
博士持ち大歓迎!