keyword
NMS-free training
NMS-free training is an object detection optimization methodology designed to eliminate the need for non-maximum suppression post-processing during inference by teaching the neural network to output exactly one distinct bounding box prediction per object. In traditional object detectors, training relies on one-to-many label assignments where multiple candidate anchors or grid cells are assigned to a single ground-truth object, necessitating non-maximum suppression at test time to filter out redundant overlapping predictions and causing additional inference latency. NMS-free training overcomes this limitation by incorporating one-to-one label matching or dual label assignment strategies, where a one-to-one prediction head is trained alongside a traditional auxiliary branch to maintain rich gradient flow during optimization. As a result, the deployed model directly outputs deduplicated predictions, enabling true end-to-end execution, lower latency, and deterministic inference on edge and real-time computing platforms without sacrificing detection accuracy.
1 item
