[paper note] Deep Attributes Driven Person Re-identification

• ECCV 2016
• paper
• poster
• Author: Chi Su, Shiliang Zhang, Junliang Xing, Wen Gao1, and Qi Tian

Intuition

• Low level feature sensitive to viewpoint, body poses, etc., and have different character corresponding to different metric learning methods.
• Attribute is mid-level feature.
• Problem: it is difficult to acquire enough training data for a large set of attributes.
• Related work: deep attribute learning works.

Three stage model

• Goal: learn an attribute predict model for person ReID through dCNN training
  $$A_I=\mathcal{O}(I)$$
• First stage: train on attribute dataset
  
  • AlexNet
  • Dataset $T=\{t_1,t_2,\dots,t_N\}$ labeled with a binary attribute.
  • Get trained model $\mathcal{O}^{S1}$, which could predict attribute for any test sample but lack discriminative power.
• Second stage: fine-tuning on person ID dataset
  
  • Dataset $U=\{u_1,u_2,\dots,u_M\}$ has person ID label.
  • Only set attributes with top $p=10$ highest confidence score as 1, others as 0.
  • Triplet loss
    $$\mathcal{L} = \sum_e^E\{\max(0, D(A^{(e)}_{(a)}, A^{(e)}_{(p)})+\theta-D(A^{(e)}_{(a)}, A^{(e)}_{(n)}))+\gamma\times \varepsilon\}$$
    $$\varepsilon=D(A^{(e)}_{(a)}, \tilde{A}^{(e)}_{(a)})+D(A^{(e)}_{(p)}, \tilde{A}^{(e)}_{(p)})+D(A^{(e)}_{(n)}, \tilde{A}^{(e)}_{(p)})$$
    where $\theta=1$, $\gamma=0.01$ in the experiment setting.
  • Get model $\mathcal{O}^{S2}$
• Third stage: fine-tuning on the Combined Dataset
  
  • First predict attributes for dataset $U$ using $\mathcal{O}^{S2}$, act as ground truth attributes.
  • Combine $T$ and $U$ as a big attribute dataset, training to get final deep attribute extractor $\mathcal{O}$

Experiment

• Attribute dataset in first stage: PETA, split multi-class attribute into binary attributes.
• Tracking dataset in second stage: MOT challenge
• Attribute accuracy
  
• Evaluate: VIPeR, PRID, GRID, Market-1501 (VIPeR, GRID and PRID are included in the PETA dataset, they will be excluded from PETA during training)
• XQDA metric learning method, further improvement
• Datasets: VIPeR (43.5%, sor->63.9%), PRID 450S (22.6%, sor->60%+), Market (Single: 39.4%, multiple: 49.0%, sor->78%)
  
  • sor denotes state-of-the-art
  • The model does not use these three dataset for training, thus it cannot compare directly with supervised re-id model.
• Additional experiments
  
  • Combine hand-crafted feature with deep attribute, improve about 8% on VIPeR.
  • Directly fine-tune FC7 feature of AlexNet, deep attribute performs better.