Skip to main content

Cascade R-CNN: Delving Into High Quality Object Detection

Citations: 1,608 | Paper Published: 18 Jun 2018 | Institutions: University of California, San Diego | Authors: Zhaowei Cai, Nuno Vasconcelos

Abstract in 2 Lines

This paper proposes a multi-stage object detection architecture called Cascade R-CNN to address the problem of overfitting and inference-time mismatch in object detection. It consists of a sequence of detectors trained with increasing IoU thresholds to be sequentially more selective against close false positives.

Contributions​

  • Proposing a new object detection architecture called Cascade R-CNN that addresses the problems of overfitting during training and inference-time mismatch between IoUs.
  • Showing that the Cascade R-CNN outperforms all single-model object detectors on the COCO dataset and is widely applicable across detector architectures.

Practical Implications​

  • The proposed Cascade R-CNN architecture can be used to improve the performance of object detection systems in various applications, such as autonomous driving, surveillance, and robotics.
  • The Cascade R-CNN can be applied to different detector architectures, making it a versatile and widely applicable approach.
  • The paper highlights the importance of addressing the problems of overfitting and inference-time mismatch in object detection, which can lead to more accurate and reliable detection systems.

Methods​

Methodology
  • Proposing a new object detection architecture called Cascade R-CNN that consists of a sequence of detectors trained with increasing IoU thresholds.
  • Training the detectors stage by stage, leveraging the observation that the output of a detector is a good distribution for training the next higher quality detector.
  • Applying the same cascade procedure at inference to enable a closer match between the hypotheses and the detector quality of each stage.

Data​

The authors have used the COCO dataset for evaluating the performance of the proposed Cascade R-CNN architecture. The COCO dataset is a widely used benchmark for object detection, segmentation, and captioning tasks. It contains a large number of images with complex scenes and a diverse set of object categories, making it a challenging dataset for object detection.

Results​

Performance Analysis
  • Cascade R-CNN with ResNet-50 backbone: achieves an AP of 42.8% at an IoU threshold of 0.5, and an AP of 23.7% at an IoU threshold of 0.9.
  • Cascade R-CNN with ResNet-101 backbone: achieves an AP of 44.2% at an IoU threshold of 0.5, and an AP of 25.1% at an IoU threshold of 0.9.

Limitations​

The limitations of this paper are not explicitly mentioned. However, one potential limitation could be that the proposed Cascade R-CNN architecture may require more computational resources and longer training time compared to single-model object detectors, due to the use of multiple stages of detectors with increasing IoU thresholds. Another limitation could be that the performance gains of the Cascade R-CNN may saturate or even degrade when applied to datasets with different characteristics or in different application domains.

Table of Contents