Exemplar-free class incremental action recognition based on self-supervised learning

Class incremental action recognition faces the persistent challenge of balancing stability and plasticity, as models must learn new classes without forgetting previously acquired knowledge. Existing methods often rely on storing original samples, which significantly increases storage demands and risks overfitting to past data. To address these issues, an exemplar-free framework based on self-supervised learning and Pseudo-Feature Generation (PFG) mechanism is proposed. At each incremental step, PFG generates pseudo features for previously learned classes by using the mean and variance for each class. This framework enables effective joint training on new class data while keeping the feature extractor frozen, eliminating the need to store original data. It preserves past knowledge and dynamically adapts to new categories, striking a balance between stability and plasticity. Experiments on four extensively used datasets: UCF101, HMDB51, Kinetics, and SSV2 validate the effectiveness of the proposed framework.