November 2019
tl;dr: Calibration of the network for a probabilistic object detector
Overall impression
The paper extends previous works in the probabilistic lidar detector and its successor. It is based on the work of Pixor.
Calibration: a probabilistic object detector should predict uncertainties that match the natural frequency of correct predictions. 90% of the predictions with 0.9 score from a calibrated detector should be correct. Humans have intuitive notion of probability in a frequentist sense. –> cf accurate uncertainty via calibrated regression and calib uncertainties in object detection.
A calibrated regression is a bit harder to interpret. P(gt < F^{-1}(p)) = p. F^{-1} = F_q is the inverse function of CDF, the quantile function.
Unreliable uncertainty estimation in object detectors can lead to wrong decision makings in autonomous driving (e.g. at planning stage).
The paper also has a very good way to visualize uncertainty in 2D object detector.
Key ideas
- This paper adds aleatoric uncertainty for each regression target based on Pixor.
- Calibration of classifiers
-
Empirical = P(label = 1 |
pred = p) = I(label = 1 and pred = p) / I (pred = p) = p = Theoretical |
- Bin prediction scores, and count empirical ones to plot the calibration plot
- Calibration of Regression
- Empirical = P(label < F_q (p)) = I(label < F_q(p)) / N = p = Theoretical
- ECE (expected calibration error) is the weighted area of calibration plot and the diagonal line, N_m is the number of samples in the m-th interval. –> see calibration of modern NN.
\[ECE = \sum_i^M \frac{N_m}{N}|p^m - \hat{p^m}|\]
- Isotonic regression (保序回归)
- During test time, the object detector produced an uncalibrated uncertainty, then corrected by the recalib model g(). In practice, we build a recalib dataset from validation data.
- Post-processing, does not guarantee recalibration of individual prediction (only by bins).
- It changes probability distribution, Gaussian –> Non-Gaussian
- Depends on the recalibration dataset.
- Temperature scaling
- One temperature per regressor (6 in total) to optimize the NLL score on recalibration dataset.
- Same distribution
- Works best with only a tiny amount of data (0.5% of entire validation dataset)
- Calibration loss
- loss between calculated $\sigma$ and regressed one
\(L_{calib} = |\sigma - (pred - gt)|\)
- Same distribution
- After recalibration, the confidence intervals are larger such that they fully cover the gt.
- Generalization to new dataset requires a tiny amount of data (~1%) to generalize (from KITTI to nuScenes).
Technical details
- Achieving higher detection does not guarantee better uncertainty estimation
- Higher detection accuracy when trained with calibration loss
Notes
- The idea of proposing one bbox per pixel/point seems to have come from the PIXOR paper. Cf LaserNet and Point RCNN.