Dense Distinct Query for End-to-End Object Detection
This work addresses a key bottleneck in end-to-end object detection for computer vision applications, offering a novel hybrid approach that improves performance in both general and crowded scenes.
The paper tackles the problem of low recall with sparse queries and optimization difficulties with dense queries in end-to-end object detection by proposing Dense Distinct Queries (DDQ), which selects distinct queries from dense ones for one-to-one assignments, achieving 52.1 AP on MS-COCO and 93.8 AP on CrowdHuman.
One-to-one label assignment in object detection has successfully obviated the need for non-maximum suppression (NMS) as postprocessing and makes the pipeline end-to-end. However, it triggers a new dilemma as the widely used sparse queries cannot guarantee a high recall, while dense queries inevitably bring more similar queries and encounter optimization difficulties. As both sparse and dense queries are problematic, then what are the expected queries in end-to-end object detection? This paper shows that the solution should be Dense Distinct Queries (DDQ). Concretely, we first lay dense queries like traditional detectors and then select distinct ones for one-to-one assignments. DDQ blends the advantages of traditional and recent end-to-end detectors and significantly improves the performance of various detectors including FCN, R-CNN, and DETRs. Most impressively, DDQ-DETR achieves 52.1 AP on MS-COCO dataset within 12 epochs using a ResNet-50 backbone, outperforming all existing detectors in the same setting. DDQ also shares the benefit of end-to-end detectors in crowded scenes and achieves 93.8 AP on CrowdHuman. We hope DDQ can inspire researchers to consider the complementarity between traditional methods and end-to-end detectors. The source code can be found at \url{https://github.com/jshilong/DDQ}.