November 2019
tl;dr: Predict bbox as two corners with mean and variance.
Similar to IoU Net, classification confidence is not always strongly related to localization confidence.
The paper models KL divergence loss between a Dirac delta label and a Gaussian prediction. Essentially this is NLL loss. For a more generalized KL loss, see LaserNet KL.
Variance voting is quite interesting idea and can be used even without variance scores (just down-weigh by IoU and weigh by confidence score). I am quite surprised this has not been tried before.
Learning localization confidence in addition to classification confidence can 1) give interpretable results 2) leads to more precise localization (AP90).
KL Loss and IoU Net are similar, but are different in implmentation. KL loss directly regresses mean and var from the same head, instead of a separate head for IoU prediction in IoU Net. Also Var Voting is one forward pass, not like the IoU Net's iterative optimization.
-
KL loss: assuming Gaussian and Dirac makes the problem more tractable and has a closed form. Same to NLL loss of a Gaussian distribution, it degenerates to a L2 loss when sigma is fixed. The authors further modifies the regression target by log transform and changed the loss to an uncertainty-aware smooth L1 loss.
$$L_{reg} = \frac{e^{-\alpha}}{2}((x_g - x_e)^2 ) + \frac{1}{2} \alpha, when <1$$ $$L_{reg} = e^{-\alpha}(|x_g - x_e| - \frac{1}{2}) + \frac{1}{2} \alpha, when >=1$$ or$$L_{reg} = e^{-\alpha} SL1(x_g - x_e) + \alpha$$ - Variance Voting: all previous NMS scheme does not weight average all bounding box but rather select one and suppress others, including IoU Guided NMS and soft NMS (only changes score). Boxes with higher variance and further away from the bbox of interest is down-weighted. The final bbox position is a weighted average of the bbox. --> this is similar to the bayesian inference scheme in BayesOD.
- The learned variance through KL loss is interpretable.
- SoftNMS can be stacked to KL loss. Var voting can also greatly boost the AP performance. It mainly comes from more accurate localization, as AP50 barely improves.
- Questions and notes on how to improve/revise the current work