Memory-based Error Label Suppression for Embodied Self-Improving Object Detection
Jieren Deng, Haojian Zhang, Jianhua Hu, Yunkuan Wang · 2024
This paper introduces a novel method for embodied self-improving object detection, aimed at enhancing the object detection model by gathering additional labeled samples post-pre-training without the need for human supervision. Current self-improving strategies autonomously label new samples using 3D consistency, yet they often incorporate a substantial amount of mislabeled samples, thereby diminishing the potential performance improvement to the model. To counter this issue, we propose a memory-based method for suppressing error labels to minimize their adverse impact. This error label suppression mechanism includes LoRA output constraint and exemplar prototype constraints, which leverage explicit memories of correct prototypes and implicit memories of correctly learned parameters, respectively. These mechanisms effectively reduce the negative effects of erroneous labels on the model’s learning process. Building upon the robustness provided by our Memory-Based Error Label Suppression, we further incorporates Cross-View Redundant Labeling to introduce a higher quantity of accurate samples, thus amplifying the benefits of embodied self-improving. Experimental results demonstrate that our method exhibits more robustness against erroneous samples compared to existing methods, leading to significantly performance improvement.