Meta Weighted Loss: Balanced Scene Graph Generation with Meta-Learning

Unbiased Scene Graph Generation SGG is a major development direction of SGG. Recent years, a number of great approaches have emerged in this field but many of them tend to overlook a fundamental factor – study on loss function. Because there is a serious conflict between datasets, loss function and pursuit metrics. In most relevant datasets, status of different predicates usually varies greatly, as the common predicate 'on' appears 400 times more frequently than the rare predicate 'walking in'. But each predicate has the same weight in the loss function which does not properly reflect their status gap in datasets. And when we evaluate results, we also treat predicates in a uniform way. Now we can sum up this conflict with an interesting statement: sometimes fairness means a kind of unfairness. In response to this challenge, we introduce Meta Weighted Loss MWL , a approach based on meta-learning. MWL leverages meta-learning principles to construct a meta-neural network during model training. This network establishes a rational relationship between various predicates and their respective weight in the loss function so that the conflict above can be solved. We verify the effectiveness and generalization of this approach on multiple datasets. Comprehensive experiments demonstrate superior performance of MWL in SGG.