Towards Calibrated Model for Long-Tailed Visual Recognition from Prior Perspective - Supplementary Material - Zhengzhuo Xu

One the one hand, the classification performance will be promoted. F ollow the settings of Proof A.2 and Eq.6, we can get the following relationship of the Hence we can generalize Eq.A.9 to: y According to Eq.A.11, it's easy to find a derivative zero point in range [1,C]. F ollow the Basic Setting of Proof.1, suppose It has limited contribution for the tail class' feature LT scenarios, the likelihood is consistent in the train and test set, but the prior is different. Hence, the actual optimization direction is not described as Eq.B.2 because the bias incurred by According to LemmaB.1, we immediately deduce that Bayias-compensated cross-entropy loss ensures All above re-weight methods are proven effective empirically, more or less. They propose the LDAM loss to encourage the tail classes to enjoy larger margins.