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
- cp=1Np∑Npif(xpi)
- Coupled cluster loss
- L(W,Xp,Xn)=∑Npi12max{0,∥f(xpi)−cp∥22+α−∥f(xn∗)−cp∥22}
- xn∗ 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 |