[Paper note] Deep Relative Distance Learning: Tell the Difference Between Similar Vehicles

• paper
• dataset

Highlight

• VGG, propose coupled clusters loss
• Multi-loss (car model, car ID)
• New dataset VehicleID

Model

• Problem with triplet loss
  
  • When margin constraint already satisfied, the training sample makes no contribution to gradient decent.
  • Ancor point sensitive (negative point is pushed away according to the selected ancor and positive point).
• Coupled clusters loss
  
  • Samples belong to the same identity should locate around a common center point.
  • Center point
    
    • $c^p = \frac{1}{N^p}\sum^{N^p}_i f(x^p_i)$
  • Coupled cluster loss
    
    • $L(W, X^p, X^n)=\sum^{N^p}_i\frac12 \max\{0, \|f(x_i^p) - c^p\|_2^2+\alpha-\|f(x^n_*)-c^p\|^2_2\}$
    • $x_*^n$ is the nearest negative sample
• Two kinds of differences: same vehicle model & same identity
  
  • 

Dataset

• VehicleID (newly proposed)

Image number	Training	Testing	All
w\ model label	47558	42638	90196
w\o model label	67540	68947	136487
All	110178	111585	226683

• CompCars
  
  • dataset
  • paper
• Tasks
  
  • Vehicle model verification
  • Vehicle retrieval
  • Vehicle re-id