I3DOL: Incremental 3D Object Learning without Catastrophic Forgetting

3D object classification has attracted appealing attentions in academic researches and industrial applications. However, most existing methods need to access the training data of past classes when facing common real-world scenario: new objects arrive a sequence. Moreover, performance advanced approaches degrades dramatically for learned (i.e., catastrophic forgetting), due irregular redundant geometric structures point cloud data. To address these challenges, we propose Incremental Object Learning I3DOL) model, which is first exploration learn continually. Specifically, an adaptive-geometric centroid module designed construct discriminative local structures, can better characterize representation object. Afterwards, prevent forgetting brought by information, geometric-aware attention mechanism developed quantify contributions explore unique characteristics with high incremental learning. Meanwhile, score fairness compensation strategy proposed further alleviate caused unbalanced between object, compensating biased prediction validation phase. Experiments on representative datasets validate superiority our I3DOL framework.