author:	DartingMelody
score:	9 / 10

What is the core idea?

The central idea of the paper is to train multiple detection heads with multiple intersection over union (IoU) thresholds. The cascade R-CNN architecture solves two problems, overfitting during training and inference-time mismatch between the IoU for which the detector is optimal versus IoU of the input hypothesis.

How is it realized (technically)?

The model consists of sequence of detectors trained with increasing IoU thresholds with the output of the previous detector fed into the next as a resampling mechanism with no discrepancy between training and inference.

• Cascade-RCNN extends the two-stage architecture of faster-RCNN relying on a cascade of specialized regressors where T is the total number of cascade stages. Each regressor \(f_{t}\) in the cascade is optimized with respect to the sample distribution {\(b^{t}\)} arriving at the t stage, instead of initial distribution \(b^{t}\) .

  

• At each stage t, the RCNN has a classifier \(h_{t}\) and regressor \(f_{t}\) which is optimized for IoU threshold \(u^{t}\), where \(u^{t}>u^{t-1}\) .

  

• This is achieved by minimizing the above loss where \(b^{t} = f_{t-1}(x^{t-1}, b^{t-1})\), g is the ground truth object for \(x^{t}\), λ = 1 is the trade-off coefficient and \(y^{t}\) is the label of \(x^{t}\) given by

  

• Cascade R-CNN has four stages, one region proposal network (RPN) and three stages for detection with U = {0.5, 0.6, 0.7}. (These were the most common IoU thresholds used by the authors unless specified.)

How well does the paper perform?

• The Cascade R-CNN, based on FPN+ and ResNet-101 outperforms all the earlier state of the art single model detectors like Faster-RCNN, YOLO, Mask-RCNN etc on the COCO dataset. It also outperforms Iterative BBox and Integral Loss models. The difference in result is more visible with higher IoU.
• As the computational cost of adding more detection head in the architecture is usually small, when compared to RPN, the computational overhead of RCNN is small for both training and testing.

What interesting variants are explored?

Various architectures like Faster-RCNN, R-FCN with ResNet-50 and ResNet-101 backbone, FPN+ with ResNet-50 and ResNet-101 backbone are trained with and without cascading. The cascade variants of all these models outperforms the corresponding non-cascade models. Other ablation experiments were performed on :

• IoU thresholds with the result that the detector can be more selective against close false positives and specialize for more precise hypothesis.
• stagewise comparison with the result that the ensemble of all classifiers is the best generally.
• regression statistics with the result that it helps the effective multi-task learning of classification and regression.
• number of stages with the result that the three stage cascade (the cascade-RCNN model ) achieves the best tradeoff.

TL;DR

• The cascade-RCNN model which extends Faster-RCNN consists of sequence of detectors trained with increasing IoU thresholds.
• The model aims to reduce overfitting, match inference and training architecture, and detect true positives while supressing close false positives.
• The Cascade-RCNN model outperforms all the previous state of the art models like Fast RCNN, YOLO, Mask-RCNN, etc on COCO dataset.