Cross-View Cross-Scene Multi-View Crowd Counting Results

  • On synthetic dataset:
  • ¬†Compared with other methods:
    • MVMS is the single-scene model
    • CamSel is the camera selection module
    • NoiseV is the noisy view regularization
  • Testing on different numbers of camera views for input:
    • the MAE decreases as more views are added.

  • On real datasets:
    • RealS – training and testing on the same real scene
    • RealC – cross-scene training on 2 real scenes, and testing on the remaining real scene.
    • Synth – cross-scene training on synthetic dataset.
    • +UDA – using unsupervised domain adaptation on the real test scene. ¬†