Class-Imbalance Aware Weighting Approach for Long-Tail Object Detection in Vision-Based IoT Systems

Addressing class imbalance is a significant issue in vision-based internet-of-things, especially when datasets have long-tail distributions, and minority classes might represent rare yet critical events, making this problem crucial. This work presents a comprehensive investigation of a weighting and reweighting approach aimed at addressing class imbalance in long-tail object detection tasks. Specifically, this study examines the effect of the weight shift factor employed in the reweighting phase, analyzes the influence of the weighting strategies on the individual loss components of the object detection model, and proposes a dynamically integrated strategy to shift the weights during model training. In addition, this work assesses the robustness and versatility of the proposed weighting and reweighting strategy by evaluating it across various datasets including aerial object detection (VisDrone-DET dataset) and ground-view long-tail detection (COCO-zipf dataset). Experimental findings utilizing the above datasets indicate that modifying the weight shift parameter proficiently regulates the intensity of weight recalculations. The analysis further indicates that the weighting strategy continues to be robust and effective, even with the dataset having a lower number of classes. Experimental results demonstrate consistent improvements across both datasets, including AP50 gains of up to 2.3 percentage points on VisDrone-DET and 9.9 percentage points on COCO-zipf compared with the corresponding baselines. Finally, deployment measurements on the NVIDIA platform indicate the viability of edge inference a potential candidate for deployment in edge analytics frameworks, including NTN-assisted IoT scenarios.