NMS Strikes Back

TL;DR

This paper compares one-to-one bipartite matching in DETRs with traditional one-to-many label assignments, finding that traditional methods with NMS outperform DETRs, and highlights the importance of architecture over matching strategy.

Contribution

It provides a comprehensive comparison showing that traditional label assignment with NMS can outperform DETRs' matching, emphasizing the role of architecture in detection performance.

Findings

Traditional NMS-based label assignment outperforms DETR matching.

Deformable-DETR achieves 50.2 COCO mAP in 12 epochs with ResNet50.

Bipartite matching is unnecessary for high-performance detection transformers.

Abstract

Detection Transformer (DETR) directly transforms queries to unique objects by using one-to-one bipartite matching during training and enables end-to-end object detection. Recently, these models have surpassed traditional detectors on COCO with undeniable elegance. However, they differ from traditional detectors in multiple designs, including model architecture and training schedules, and thus the effectiveness of one-to-one matching is not fully understood. In this work, we conduct a strict comparison between the one-to-one Hungarian matching in DETRs and the one-to-many label assignments in traditional detectors with non-maximum supervision (NMS). Surprisingly, we observe one-to-many assignments with NMS consistently outperform standard one-to-one matching under the same setting, with a significant gain of up to 2.5 mAP. Our detector that trains Deformable-DETR with traditional IoU-based label assignment achieved 50.2 COCO mAP within 12 epochs (1x schedule) with ResNet50 backbone, outperforming all existing traditional or transformer-based detectors in this setting. On multiple datasets, schedules, and architectures, we consistently show bipartite matching is unnecessary for performant detection transformers. Furthermore, we attribute the success of detection transformers to their expressive transformer architecture. Code is available at https://github.com/jozhang97/DETA.