Generalized Focal Loss

发表于 |

Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection

  1. 动机

    • one-stage detectors
      • dense prediction
      • three fundamental elements
        • class branch
        • box localization branch
        • an individual quality branch to estimate the quality of localization
    • current problems
      • the inconsistent usage of the quality estimation in train & test
      • the inflexible Dirac delta distribution：将box regression的value建模成真值附近的脉冲分布，用来描述边界不清晰/遮挡的case可能不准确
    • we design new representations for these three elements
      • merge quality estimation into class prediction：将objectness/centerness整合进cls prediction，直接用作NMS score
      • continout labels
      • propose GFL(Generalized Focal Loss) that generalizes Focal Loss from discrete form into continous version
    • test on COCO
      • ResNet-101-?-GFL: 45.0% AP
      • defeat ATSS

  2. 论点

    • inconsistent usage of localization quality estimation and classification score

      • 训练的时候quality和cls branch是independent branch
      • box branch的supervision只作用在positive样本上：which is unreliable on predicting negatives
      • 测试阶段将quality和cls score乘起来有可能拉高负样本的分数，以至于在NMS阶段把低分正样本挤掉

    • inflexible representation of bounding boxes

      • most method建模成脉冲分布：只在IoU大于一定阈值的格子上有响应，别的格子都是0
      • some recent work建模成高斯分布
      • in fact the real distribution can be more arbitrary and flexible，连续且不严格镜像

    • thus we propose

      • merge the quality representation into the class branch：

        • class vector的每个元素代表了格子的localization quality(如IoU score)

        • 在inference阶段也是直接用作cls score

      • propose arbitrary/general distribution

        • 有明确边界的目标的边的分布是比较sharp的

        • 没有明显边界的边分布就是flatten一点

      • Generalized Focal Loss (GFL)

        • joint class representation是continuous IoU label (0∼1)
        • imbalance问题仍然存在，但是standart Focal Loss仅支持[0,1] sample
        • 修改成continuous形式，同时specialized into Quality Focal Loss (QFL) and Distribution Focal Loss (DFL)
          • QFL for cls branch：focuses on a sparse set of hard examples
          • DFL for box branch： focus on learning the probabilities of values around the continuous target locations

  3. 方法

    • Focal Loss (FL)

      • standard CE part：$-log(p_t)$
      • scaling factor：down-weights the easy examples，focus on hard examples

    • Quality Focal Loss (QFL)

      • soft one-hot label：正样本在对应类别上有个(0,1]的float score，负样本全0

      • float score定义为预测框和gt box的IoU score

      • we adopt multiple binary classification with sigmoid

      • modify FL

        • CE part 改成complete form：$-ylog(\hat y)-(1-y)log(1-\hat y)$
        • scaling part用vector distance替换减法：$|y-\hat y |^{\beta}$

        • $\beta$ controls the down-weighting rate smoothly & $\beta=2$ works best

    • Distribution Focal Loss (DFL)

      • use relative offsets from the location to the four sides of a bounding box as the regression targets

      • 回归问题formulation

        • 连续：$\hat y = \int_{y_0}^{y_n}P(x)xdx$
        • 离散化：$\hat y = \sum_{i=0}^n P(y_i)y_i$
        • P(x) can be easily implemented through a softmax layer containing n+1 units：

      • DFL

        • force predictions to focus values near label $y$：explicitly enlarge the probabilities of $y_i$和$y_{i+1}$，given $y_i \leq y \leq y_{i+1}$
        • $log(S_i)$ force the probabilities

        • gap balance the 上下限，使得$\hat y$的global mininum solution无限逼近真值$y$，如果真值接近的是$\hat y_{i+1}$，可以看到$log(S_i)$那项被downscale了

    • Generalized Focal Loss (GFL)

      • 以前的cls preditions在测试阶段要结合quality predictions作为NMS score，现在直接就是

      • 以前regression targets每个回归一个值，现在是n+1个值

      • overall

        • 第一项cls loss，就是QFL，dense on 所有格子，用正样本数去norm
        • 第二项box loss，GIoU loss + DFL，$\lambda_0$默认2，$\lambda_1$默认1/4，只计算有IoU的格子
        • we also utilize the quality scores to weight $L_B$ and $L_D$ during training

  4. 彩蛋

    • IoU branch always superior than centerness-branch

      • centerness天生值较小，影响召回，IoU的值较大