Comparative Evaluation Using Backbone Architecture Cascade Mask R-CNN for Robust Waste Instance Segmentations

Environmental pollution from inorganic waste, particularly persistent plastics and metals, poses escalating threats to global ecosystems and public health. Existing waste detection systems face critical accuracy limitations in real-world environments due to fragmented waste distribution and complex backgrounds. With Indonesia generating over 18 million tons of waste annually and recycling rates below 40 %, automated monitoring solutions are urgently needed to support sustainable waste management. We evaluate three backbones-ResNet-101, ResNeXt-101, and ConvNeXt-in Cascade Mask R-CNN to optimize waste segmentation, achieving $0.597 \text{mAP} {@} 50$ detection accuracy and $0.486 \text{mAP} {@} 75$ segmentation precision using ConvNeXt. The study demonstrates particular effectiveness for container-level waste ($0.664 \text{mAP} \_\mathrm{l}$) but reveals ongoing challenges with small debris (0.125 mAP _s) and irregular residual waste (0.185 mAP). Future work will explore expanded datasets, including industrial waste scenarios, and integrate real-time monitoring for smart city deployment.