PseDet: Revisiting the Power of Pseudo Label in Incremental Object Detection

Task Incremental Object Detection (IOD) aims to overcome catastrophic forgetting while expanding the detector's ability to recognize new classes incrementally. Motivation • Quality Limitation of Pseudo Labels: The quality of pseudo labels generated by the previous model is constrained by the teacher model's performance, which may introduce noisy data and degrade the learning. • Confidence biases across different categories: Using a fixed threshold for filtering pseudo labels across all classes ignores the varying score distributions among different categories. • Misalignment of Confidence Scores and Localization Quality: The confidence scores do not linearly correlate with the localization quality of pseudo labels, which can introduce noise if directly used for training. Categorical Adaptive Label Selector It dynamically determines the class-wise filtering threshold with a simple mathematical prior and fast K-Means pre-computation. Methods Experiments One-Step Results Multi-Step Results Ablations Methods The overall framework of PseDet for incremental object detection: Spatio-Temporal Enhancement The Module alleviate the negative effects when learning noisy data from the previous model by reducing spatial noise through multi-scale augmentation and fusion, and mitigating temporal noise accumulation across incremental steps. Confidence Score Calibration Supervision It calibrates the distribution of confidence in order to align the score with the localization quality of the pseudo labels, then integrates the quality into the new-step supervision.