Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces an Online Active Learning (OAL) framework that enables compact neural networks to adapt in real-time to dynamic environments by continuously integrating temporal visual cues from monocular vision streams. By employing a Predict-Evaluate-Correct learning paradigm alongside Elastic Weight Consolidation (EWC), the framework allows for selective plasticity, preserving critical structural knowledge while adapting to new data. The approach achieves a 75% reduction in computational cost while maintaining competitive depth estimation accuracy, highlighting that adaptability is more about effective parameter regulation than model size.
Real-time depth perception can be achieved with a 75% reduction in computational cost by leveraging selective plasticity in compact neural networks.
Biological visual systems can perceive depth from monocular vision flow, continuously integrating temporal visual cues while maintaining a balance between stability and plasticity in dynamic environments. In contrast, artificial perception models deployed on resource-constrained edge devices are typically trained in a static offline manner and remain frozen after deployment, often suffering severe performance degradation under domain shifts. While large-scale models may encode broad knowledge through massive parameter redundancy, lightweight networks face a static optimization dilemma: forcing compact models to learn universal geometric representations is computationally inefficient and often leads to performance saturation. To resolve this issue, an Online Active Learning (OAL) mechanism is introduced to endow compact neural networks with the capability to adapt continuously during operation. A closed-loop Predict-Evaluate-Correct learning paradigm is established to actively select high-confidence, information-rich signals from streaming visual input. Crucially, Elastic Weight Consolidation (EWC) is employed not merely to prevent catastrophic forgetting, but to enforce Selective Plasticity, preserving parameters that encode globally relevant structural knowledge while allowing local alignment to newly observed environments. Built upon a MobileNetV3-Small backbone, the proposed system achieves approximately a 75% reduction in computational cost while maintaining competitive depth estimation accuracy. Experimental results demonstrate that adaptability is not solely determined by model size, but rather by how effectively parameter plasticity is regulated in dynamic environments.