# Cascade-DETR: Delving into High-Quality Universal Object Detection

**Authors:** Mingqiao Ye, Lei Ke, Siyuan Li, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, Fisher Yu  
**arXiv:** [2307.11035](https://arxiv.org/abs/2307.11035)  
**Date:** 2023-07-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.11035v1/

## Key points

- Cascade-DETR is a novel approach for achieving high-quality universal object detection
- It introduces the Object-Centric Transformer layer to improve generalization and localization accuracy
- Cascade-DETR predicts expected IoU instead of relying on classification scores, leading to well-calibrated confidences
- A new benchmark called UDB10 has been introduced to evaluate performance, showing significant improvements over DETR-based detectors
- Authors Mingqiao Ye, Lei Ke, Siyuan Li, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, and Fisher Yu have contributed to this research accepted at ICCV 2023

## Summary

, , , , 

Cascade-DETR: A Novel Approach to High-Quality Universal Object Detection

 is a groundbreaking method for achieving high-quality universal object detection. It addresses the challenges faced by recent Transformer-based detection methods in diverse domains, where accurately localizing objects in complex scenarios is crucial for vision systems. One of the key innovations of Cascade-DETR is the introduction of the . This layer integrates object-centric information into the detection decoder by restricting attention to previous box predictions. This not only improves generalization to diverse domains but also enhances localization accuracy. In addition, Cascade-DETR revisits the scoring of queries by predicting the expected  instead of relying on classification scores. This leads to more well-calibrated confidences and further enhances accuracy in object detection tasks. To evaluate its performance, a new universal object detection benchmark called  has been introduced. It consists of 10 datasets from various domains and Cascade-DETR not only advances the state-of-the-art on COCO but also substantially improves DETR-based detectors on all datasets within UDB10, with improvements exceeding 10 mAP in some cases. These enhancements are particularly notable under stringent quality requirements, showcasing the effectiveness of Cascade-DETR in achieving high-quality universal object detection. The authors Mingqiao Ye, Lei Ke, Siyuan Li, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, and Fisher Yu have contributed to this research accepted at ICCV 2023. The code and models for Cascade-DETR will be made available at https://github.com/SysCV/cascade-detr for further exploration and implementation in vision systems.