Efficient Parameter Mining and Freezing for Continual Object Detection
Angelo G. Menezes, Augusto J. Peterlevitz, Mateus A. Chinelatto, André C. P. L. F. de Carvalho
TL;DR
This work tackles catastrophic forgetting in continual object detection (COD) by introducing layer-level parameter mining and freezing. It defines four layer-importance criteria based on internal feature-map statistics (mean, median, variance, information entropy) and explores a gradient-penalty relaxation to inject limited plasticity; evaluations are conducted on incremental Pascal VOC and the TAESA dataset. The results show that entropy-based layer freezing often yields the best stability-plasticity tradeoffs and can outperform neuron-level mining in some settings, but generally remains behind knowledge-distillation and replay baselines; gradient penalties offer limited gains and a hybrid approach with replay may be most practical. Overall, the study highlights layer-wise parameter isolation as a scalable strategy for COD with clear directions for future refinements, including finer-grained freezing and task-relatedness-aware coefficients.
Abstract
Continual Object Detection is essential for enabling intelligent agents to interact proactively with humans in real-world settings. While parameter-isolation strategies have been extensively explored in the context of continual learning for classification, they have yet to be fully harnessed for incremental object detection scenarios. Drawing inspiration from prior research that focused on mining individual neuron responses and integrating insights from recent developments in neural pruning, we proposed efficient ways to identify which layers are the most important for a network to maintain the performance of a detector across sequential updates. The presented findings highlight the substantial advantages of layer-level parameter isolation in facilitating incremental learning within object detection models, offering promising avenues for future research and application in real-world scenarios.