Compact Minutiae Detection Model using a Non-NMS Tiny Object Detection
We explore a novel strategy to reduce redundant predictions in object detection by aligning the training and inference stages more closely. Traditional detectors often rely on a one-to-many assignment strategy during training, which leads to excessive outputs and necessitates post-processing such as non-maximum suppression (NMS). Our approach addresses this limitation by incorporating a classification-aware objective that naturally encourages one-to-one correspondence between predictions and ground-truth objects. This allows the model to eliminate the need for NMS and perform end-to-end detection more effectively.