Cascade R-CNN: High Quality Object Detection and Instance Segmentation (1906.09756v1)

Abstract: In object detection, the intersection over union (IoU) threshold is frequently used to define positives/negatives. The threshold used to train a detector defines its \textit{quality}. While the commonly used threshold of 0.5 leads to noisy (low-quality) detections, detection performance frequently degrades for larger thresholds. This paradox of high-quality detection has two causes: 1) overfitting, due to vanishing positive samples for large thresholds, and 2) inference-time quality mismatch between detector and test hypotheses. A multi-stage object detection architecture, the Cascade R-CNN, composed of a sequence of detectors trained with increasing IoU thresholds, is proposed to address these problems. The detectors are trained sequentially, using the output of a detector as training set for the next. This resampling progressively improves hypotheses quality, guaranteeing a positive training set of equivalent size for all detectors and minimizing overfitting. The same cascade is applied at inference, to eliminate quality mismatches between hypotheses and detectors. An implementation of the Cascade R-CNN without bells or whistles achieves state-of-the-art performance on the COCO dataset, and significantly improves high-quality detection on generic and specific object detection datasets, including VOC, KITTI, CityPerson, and WiderFace. Finally, the Cascade R-CNN is generalized to instance segmentation, with nontrivial improvements over the Mask R-CNN. To facilitate future research, two implementations are made available at \url{https://github.com/zhaoweicai/cascade-rcnn} (Caffe) and \url{https://github.com/zhaoweicai/Detectron-Cascade-RCNN} (Detectron).

• The paper introduces a multi-stage detector that refines object hypotheses using progressively higher IoU thresholds.
• It applies cascaded bounding box regression and detection heads to improve localization precision and classification accuracy.
• Evaluations on COCO and other benchmarks demonstrate state-of-the-art performance in both detection and instance segmentation.

Cascade R-CNN: High Quality Object Detection and Instance Segmentation

The paper "Cascade R-CNN: High Quality Object Detection and Instance Segmentation" by Zhaowei Cai and Nuno Vasconcelos presents a multi-stage object detection architecture designed to address the challenges of high-quality object detection.

Overview

Object detection is a complex problem involving two main tasks: recognition and localization. The standard Intersection over Union (IoU) threshold of 0.5 commonly used in training object detectors yields noisy detections and performance degradation at higher thresholds. This issue is attributed to the paradox of high-quality detection, which includes overfitting due to a vanishing number of positive samples at high IoU thresholds and a mismatch between training and inference-time quality. The Cascade R-CNN is proposed to resolve these challenges by employing a sequence of detectors trained with progressively increasing IoU thresholds, thus enhancing the quality of hypotheses and ensuring consistency between training and inference.

Architecture

The Cascade R-CNN utilizes a multi-stage extension of the R-CNN framework. Each stage consists of a detection head that includes a classifier and a regressor, trained with increasing IoU thresholds. The key aspects of this architecture are:

1. Cascaded Bounding Box Regression: Each stage improves the hypotheses sequentially, ensuring high-quality hypotheses at each stage and minimizing overfitting by maintaining a large number of positive examples.
2. Cascaded Detection: By training stages sequentially with resampled higher quality hypotheses, the same cascade applied during inference ensures a match between the quality of the detector and the hypotheses.

Comparative Analysis

The paper also compares the Cascade R-CNN with two primary techniques: iterative bounding box (BBox) regression and integral loss.

• Iterative BBox Regression: This method applies the same regressor iteratively, leading to suboptimal results due to a mismatch between the training distribution of the regressor and the higher quality of iteratively refined proposals.
• Integral Loss: While this method learns multiple classifiers for different IoU thresholds, it does not solve the imbalance in positive samples across thresholds or the proposal quality mismatch during inference.

Evaluation

The proposed Cascade R-CNN demonstrates state-of-the-art performance across multiple robust datasets and various baseline network architectures. The method shows consistent improvements in AP, particularly at high IoU thresholds, thus validating its effectiveness in high-quality object detection.

• COCO: Significant improvements are reported on the COCO dataset, with the Cascade R-CNN outperforming competitive models like Faster R-CNN and Mask R-CNN by notable margins, especially at higher IoU thresholds.
• Generalization: The Cascade R-CNN shows robust performance gains on PASCAL VOC, KITTI, CityPersons, and WiderFace datasets, irrespective of the baseline detector architecture used.
• Instance Segmentation: The extension to instance segmentation, referred to as Cascade Mask R-CNN, also results in performance gains over Mask R-CNN. This approach allows consistent advances in both the detection and segmentation tasks due to the utilization of high-quality proposals.

Practical Implications and Future Work

The Cascade R-CNN's ability to leverage multiple stages to progressively improve hypotheses and detectors is not only computationally efficient but also broadly applicable across different detection architectures and tasks. The method's robustness introduces a valuable paradigm shift for object detection, particularly for applications that require precise localization, such as autonomous driving and face detection.

Conclusion

The Cascade R-CNN submitted by Zhaowei Cai and Nuno Vasconcelos establishes a new framework for handling high-quality object detection and instance segmentation. By addressing the paradox of high-quality detection through a multi-stage resampling and training mechanism, the Cascade R-CNN provides substantial accuracy and robustness improvements. This system stands to be an essential tool in future object detection research and practical applications, facilitating enhanced performance metrics and ensuring reliable high-quality detection outcomes.