Figures (10)  Tables (1)

    • Figure 1. 

      Sample images of different vehicle models from (a) the VehicleID and (b) VeRi-776 datasets.

    • Figure 2. 

      The pros and cons of the VehicleID dataset: (a) visible faces, (b) noisy red time stamp, (c) only two directions. (The images were captured in high quality).

    • Figure 3. 

      CycleGAN model architecture transferring between VeRi-776 (real domain) and VehicleX (synthetic domain).

    • Figure 4. 

      Left: faces are blurred out, such that the privacy of the driver is ensured; Right: the timestamps in red are (partially) erased.

    • Figure 5. 

      Generated views by CycleGAN.

    • Figure 6. 

      Left: background removal; Right: cropped vehicles still look like vehicles.

    • Figure 7. 

      I spy with my convoluted eyes: (a) a taxi, (b) a bush, (c) an abstract self portrait, (d) a fence.

    • Figure 8. 

      VeRi-776: samples of (a) taxis and yellow fence, and (b) vehicles captured behind bushes and non-blackened out license plates.

    • Figure 9. 

      Comparing the two experiments: the outcomes of B are qualitatively more reliable than A.

    • Figure 10. 

      Generator loss, discriminator loss and forward cycle consistency loss curves per epoch for VCGAN-A and VCGAN-B.

    • VehicleID VeRi-776
      VCGAN-A 37,778 37,778
      VCGAN-B 113,346 37,778

      Table 1. 

      Training data sizes for VCGAN-A and VCGAN-B.