Recognizing rear-seat occupancy is an essential aspect of the intelligent passenger cabin environment since seatbelt reminder, restraints supervision, child presence, and post-collision seat occupant counting depend on accurate information about the status of left, center, and right seats. Millimeter-wave radars are suitable sensors to use in this case due to non-contact and privacy-protected measurements in the absence of illumination and partial visual occlusion. This paper contributes a topology-specific approach to rear-row occupancy recognition with a radar by considering the cabin as a fixed three-seat decision surface rather than an open object detection scene. SAML-Net architecture includes a compact shared encoder, differentiable seat prior pooling, three seatwise occupancy heads, an auxiliary eight-state classifier, and count consistency loss. This manuscript extends the literature review, explains the conversion from signal data to heatmaps, describes each mathematical component of the model, and connects the reported experimental findings to the actual nature of the problem. Frame-level tensors of radar data and full annotation files have not been released, which makes it impossible to verify training process. This paper addresses the core research question of the experiment without making any unsubstantiated claims about training at the level provided by the experiment: multi-label classification based on a fixed topology is a better representation of rear-seat occupancy problem for the radar heatmaps than proposal-based detection.