Relaxed Softmax: Efficient Confidence Auto-Calibration for Safe Pedestrian Detection

As machine learning moves from the lab into the real world, reliability is often of paramount importance. The clearest example are safety-critical applications such as pedestrian detection in autonomous driving. Since algorithms can never be expected to be perfect in all cases, managing reliability becomes crucial. To this end, in this paper we investigate the problem of learning in an end-to-end manner object detectors that are accurate while providing an unbiased estimate of the reliablity of their own predictions. We do so by proposing a modification of the standard softmax layer where a probabilistic confidence score is explicitly pre-multiplied into the incoming activations to modulate confidence. We adopt a rigorous assessment protocol based on reliability diagrams to evaluate the quality of the resulting calibration and show excellent results in pedestrian detection on two challenging public benchmarks.