Class Rectification Hard Mining for Imbalanced Deep Learning
Citations
1,447 citations
Cites background from "Class Rectification Hard Mining for..."
...Another line of important work aims to study sample difficulty in terms of loss and assign higher weights to hard examples [11, 29, 8, 26]....
[...]
780 citations
Cites methods from "Class Rectification Hard Mining for..."
...Some recent methods include metric learning [22, 33], hard negative mining [10, 27], and meta learning [15, 55]....
[...]
684 citations
Cites background from "Class Rectification Hard Mining for..."
...Among the previous works, the one presented in [9] is the most relevant to ours....
[...]
...Our margin ranking loss differs from [9] in the mining of the positives and hard negatives which is more efficient and specialized for incremental learning....
[...]
...[9] propose a class rectification loss to rectify the learning bias of cross-entropy loss given the imbalanced data....
[...]
...Previous efforts to tackle the class imbalance can be roughly divided into two groups: data resampling [13, 6, 12] and cost-sensitive learning [32, 17, 38, 16, 9]....
[...]
331 citations
Cites methods from "Class Rectification Hard Mining for..."
...The idea of reweighting examples can be dated back to dataset resampling [40, 41] or instance re-weight [42], which pre-evaluates the sample weights as a preprocessing step by using certain prior knowledge on the task or data....
[...]
313 citations
References
73,978 citations
49,914 citations
30,843 citations
"Class Rectification Hard Mining for..." refers methods in this paper
...The CRL approach shares a similar principle to Batch Normalisation [23] for easing network optimisation....
[...]
30,811 citations
30,462 citations
"Class Rectification Hard Mining for..." refers background in this paper
...Datasets ILSVRC2012-14 [37] COCO [29] VOC2012 [12] CIFAR-100 [26] Caltech 256 [18] CelebA [32] DeepFashion [31] X-Domain [7]...
[...]
...For COCO [29], no specific numbers are available for calculating between-class imbalance ratios, mainly because the COCO images often contain simultaneously multiple classes of objects and also multiple instances of a specific class....
[...]