November 2019
tl;dr: Model aleatoric uncertainty in both RPN and FRN.
This work extends previous work towards safe ad.
After modeling aleatoric uncertainty, performance is boosted almost 9%.
As we are modeling aleatoric uncertainty for each bbox, it is hetereoscedastic by nature.
However they did not report how prediction quality changes with predicted uncertainty.
- RPN generates 3D bbox that are axis-aligned. Then the RoIPooled features then regress four corners in BEV, top and bottom surface position, and encoded orientation (sin, cos).
- Aleatoric classification uncertainty is not explicitly modeled as it is self-contained from the softmax score. All regression loss are in the form of uncertainty aware loss.
$$L_{uncertainty} = e^{-t} L + t$$ - Train without uncertainty until almost converges, then add uncertainty. This trains faster.
- Modeling in FRH gives best performance in easy setting, and in both RPN and FRH gives best performance in moderate and hard.
- Use TV (total variance) to quantify aleatoric uncertainty.
$TV = \sum_i \sigma_i^2$ - Uncertainty findings:
- The uncertainty grows with off-base axis angles
- Uncertainty distribution has peak at smaller value compared to hard setting.
- Uncertainty decreases with increasing softmax score --> This is a bit contradictory to previous findings. Maybe modeling uncertainty made softmax scores more representative to indicate uncertainty?
- Summary of technical details
- Focal loss and online hard negative mining focuses on hard examples. However modeling uncertainty ignores noisy (and potentially hard) examples. How does a neural net distinguish this is a hard example or a noisy one?