Improper segregation of biomedical waste allows infectious and otherwise hazardous material to enter general waste streams, with consequences for public health and the environment. Reliable automatic classification at the point of generation could support safer segregation and disposal. Deep learning has been applied to this task, but the public image datasets used for training and evaluation are commonly distributed in a pre-augmented form, in which each source photograph appears as many transformed copies. A conventional random train–test split then places near-identical copies in both sets, so the reported accuracy partly reflects memorisation rather than generalisation. We study automatic classification of 23 biomedical waste categories (15,600 background-removed images) with explicit attention to this issue. Because the dataset carries no ground-truth labelling of duplicates, a leakage-free partition cannot be certified; instead, visually near-identical images are grouped from image content using ORB keypoints with RANSAC geometric verification and kept within a single split, and this controlled estimate is compared against the fully mixed random split, which is known to be leaked, as an optimistic reference. Across four transfer-learning backbones, ConvNeXt-Tiny performs best, reaching a macro-F1 of 0.847±0.007 under the controlled split against 0.958±0.003 under the random split, so naive evaluation is optimistic by about eleven points. Mapping the categories onto World Health Organization hazard streams yields a disposal-bin accuracy of 0.892, with 1.76% of hazardous items misrouted to general waste. The study offers a benchmark and a more realistic basis for reporting automatic biomedical-waste classification.