On Improving Convolutional Networks Based People Detection with Fisheye Cameras

In this paper, we propose a new method to train convolutional neural networks for detecting people in images taken with ceiling-mounted fisheye cameras. While simply fine-tune existing detectors using annotated images lead to increased false positives due to lack of variety in the training data, we find that adding automatically computed backgrounds of the target scene in the training process yields much better detection accuracies. This allows us to build practical scene-specific human detectors.