RobustEWaste: A Data-Centric Robustness Evaluation of Lightweight E-Waste Detection Under Simulated Smart-Bin Conditions

Electronic waste is the stream of discarded electrical and electronic equipment, and e-waste detection automates its sorting by having a camera recognize items such as phones, batteries, and circuit boards. This paper measures how much of a lightweight detector’s accuracy survives the imaging conditions a real smart bin would impose. Published e-waste detectors report strong accuracy, but almost all are trained and tested on a single dataset under uniform conditions, so their reported figures say little about performance under the poor lighting, occlusion, and compression a deployed camera actually encounters. We present RobustEWaste, a data-centric evaluation framework that harmonizes inconsistent public label vocabularies into an 8-class taxonomy, audits the result for duplicates and annotation defects, applies a reproducible software corruption pipeline covering low brightness, glare, shadow, motion blur, compression, rotation, and occlusion, and reports a Robustness Drop metric alongside standard accuracy. On YOLO11s, smart-bin augmentation with class-balanced oversampling raised clean mAP@0.5 from 0.723 to 0.830 and corrupted mAP@0.5 from 0.296 to 0.356 at no inference cost, running near 90 FPS on a T4 GPU. Robustness Drop nonetheless stayed above 57 percent for both configurations, and the errors were overwhelmingly missed detections rather than class confusions. Clean-set accuracy therefore overstates deployable performance by more than half.