Data-Efficient Computer Vision for Environmental Monitoring in Resource-Constrained Settings

Abstract

Computer vision can expand the scope of garbage detection and reduce the reliance on continuous manual observation, and is expected to be applied in a wider range of environmental monitoring. However, these methods still face some problems in practical applications, such as limited labeled data, significant differences in the number of samples for different categories, and limited computing power of edge devices. This study conducted experiments based on the public Trash Annotations in Context dataset. This dataset contains 1500 outdoor images and 4784 garbage instances. The research mainly focused on how to maintain good detection performance with limited labeled data and device resources. To this end, the study adopted uncertainty-diversity active selection, category-balanced focal training, and a lightweight SSDLite–MobileNetV3 model, and conducted tests under 5%, 10%, 20%, 40%, and 100% of the labeled data. When only using 20% of the labeled data, the mAP@[0.5:0.95] of the complete method reached 0.319, equivalent to 89.4% of the performance of the fully supervised SSDLite benchmark model. When the labeled data increased to 40%, this proportion further improved to 97.8%. Category-balanced training also improved the detection performance of minority categories, with the average precision increasing from 0.167 to 0.224 and the macro recall rate increasing from 0.488 to 0.568. In terms of device deployment, the latency of the INT8 model was reduced by 33.8% compared to the FP32 model, the model size was reduced by 70.5%, the peak memory usage decreased by 27.4%, at the cost of a 1.2 percentage point decrease in mAP. In summary, this method can maintain good detection performance with less labeled data and limited device resources, but there is still a certain trade-off between accuracy and deployment cost, and whether it can be applied to other environmental monitoring tasks still needs further verification.