author:	mmcinnestaylor
score:	8 / 10

Core Idea

• In object detection, an intersection over union (IoU) threshold is used to define the percentage of overlap of a predicted bounding box over the ground truth bounding box in order to be considered a positive sample.
• Cascade R-CNN uses a sequence of detectors with increasing IoU thresholds, leading to higher selectivity against false positives. The output distribution of a lower quality detector at the previous stage is used as input for a higher quality detector at the successive stage.
• The cascade procedure helps reduce overfitting during training, and leads to closer matches between the hypothesis and detector quality at each stage during inference.

Technical Implementation

• The architecture is framed as a cascaded regression problem:
  • \[f(x,\textbf{b}) = f_T\circ f_{T-1}\circ...\circ f_1(x,\textbf{b})\]
    • Here, \(T\) is the total number of cascade stages.
  • Each regressor \(f_t\) is optimized w.r.t. the sample distribution \(\textbf{b}^t\) which arrives at stage \(t\) instead of the initial distribution \(\textbf{b}^1\)

• Each stage \(t\in T\) includes a classifier \(h_t\) and regressor \(f_t\) which are optimized for an IoU threshold \(u^t\) where \(u^t>u^{t-1}\)
• The above is achieved by minimizing the following loss:
  • \[L(x^t, g) = L_{cls}(h_t(x^t),y^t) + \lambda[y^t\geq 1]L_{loc}(f_t(x^t,\textbf{b}^t),\textbf{g})\]
    • \[\textbf{b}^t = f_{t-1}(x^{t-1},\textbf{b}^{t-1})\]
    • And where \(g\) is the ground truth object for \(x^t\), \(\lambda = 1\) is the tradeoff coefficient, \([\cdot]\) is the indicator function, and \(y^t\) is the label of \(x^t\) given \(u^t\)
  • The function is the sum of two terms related to seperate tasks in object detection.
  • The first term in the loss is related to the classification task, while the second term is related to the bounding box regression task
  • The tradeoff coefficient helps manage the contribution of the two tasks to the overall loss

Variants

• Experiments were performed implementing the architecture while varying the number of cascade stages. The results are summarized below.
  • 1~n denotes an emsemble of n classifiers on proposals at a stage

Results

• The first table compares results of multiple state of the art single model object detectors.
  • The first block are one-stage, the second block are two-stage, and the third block multi-stage
  • Cascade R-CNN outperformed all other detectors in this comparison

• The second table summarizes the cascade procedure applied to different detection architectures. The consistent improvement in detection performance led the authors to suggest that Cascade R-CNN had strong generalization capacity.
• The authors note that since the computational cost of a detection head is usually small compared to the Region Proposal Network, the computational overhead of the cascade procedure is thus small both during training and testing.

TL;DR

• Cascade R-CNN uses a series of connected stages with increasing IoU thresholds to achieve state of the art detection performance
• Experiments demonstrate the detection performance comes as a small increase in computational cost
• Experiments also demonstrate the architecture to be effective across a number of different backbones