Real-time object recognition on edge devices with constrained computing resources involves a trade-off between computational workload and classification accuracy. Existing classifier models are individual classifiers that are typically designed for either fast inference with reduced accuracy or high accuracy with significant computational cost. A recently proposed concept, called IDK (which stands for "I don’t know") classifiers, enables cascading multiple existing classifiers to achieve high accuracy while significantly reducing average inference time. In this work, we compose IDK classifier cascades for the Tiny ImageNet dataset. Each input is processed sequentially through classifiers—from faster, less accurate ones to slower, more accurate ones. When an upstream classifier returns a high-confidence prediction, downstream models are skipped, improving average inference time. Our experiments demonstrate that IDK classifier cascades can reduce average computation time per inference while maintaining high classification accuracy compared to state-of-the-art individual models.