guided anchoring

发表于 |

原作者知乎reference：https://zhuanlan.zhihu.com/p/55854246

  • 不完全是anchor-free，因为还是有decision grid to choose from的，应该说是adaptive anchor instead of hand-picked
  • 为了特征和adaptive anchor对齐，引入deformable conv

Region Proposal by Guided Anchoring

  1. 动机

    • most methods
      • predefined anchors
      • do a uniformed dense prediction
    • our method
      • use sematic features to guide the anchoring
      • anchor size也是网络预测参数，compute from feature map
      • arbitrary aspect ratios
    • feature inconsistency
      • 不同的anchor loc都是对应feature map上某一个点
      • 变化的anchor size和固定的位置向量之间存在inconsistency
      • 引入feature adaption module
    • use high-quality proposals
      • GA-RPN提升了proposal的质量
      • 因此我们对proposal进入stage2的条件更严格
    • adopt in Fast R-CNN, Faster R-CNN and RetinaNet均涨点
      • RPN提升显著：9.1
      • MAP也有涨点：1.2-2.7
    • 还可以boosting trained models
      • boosting a two-stage detector by a fine-tuning schedule

  2. 论点

    • alignment & consistency

      • 我们用feature map的pixels作为anchor representations，那么anchor centers必须跟feature pixels保持align

      • 不同pixel的reception field必须跟对应的anchor size保持匹配

      • previous sliding window scheme对每个pixel都做一样的操作，用同样一组anchor，因此是align和consist的
      • previous progressly refining scheme对anchor的位置大小做了refinement，ignore the alignment & consistency issue，是不对的！！

    • disadvantage of predefined anchors

      • hard hyperparams
      • huge pos/neg imbalance & computation

    • we propose GA-RPN

      • learnable anchor shapes to mitigate the hand-picked issue
      • feature adaptation to solve the consistency issue
      • key concerns in this paper
        • learnable anchors
        • joint anchor distribution
        • alignment & consistency
        • high-quality proposals

  3. 方法

    • formulation

      • $p(x,y,w,h|I) = p(x,y|I)p(w,h|x,y,I)$
      • 将问题解耦成位置和尺寸的预测，首先anchor的loc服从full image的均匀分布，anchor的size建立在loc存在的基础上
      • two branches for loc & shape prediction
        • loc：binary classification，hxwx1
        • shape：location-dependent shapes，hxwx2
        • anchors：loc probabilities above a certain threshold & correponding ‘most probable’ anchor shape
      • multi-scale
        • the anchor generation parameters are shared

      • feature adaptation module

        • adapts the feature according to the anchor shape

    • anchor location prediction

      • indicates the probability of an object’s center
      • 一层卷积：1x1 conv，channel1，sigmoid
      • transform back：each grid(i,j) corresponds to coords ((i+0.5)*s, (j+0.5)*s) on the origin map
      • filter out 90% of the regions
      • thus replace the ensuing conv layers by masked convs
      • groud truth
        • binary label map
        • each level：center region & ignore region & outside region，基于object center的方框
          • $\sigma_1=0.2，\sigma_2=0.5$：region box的长宽系数
          • ？？？用centerNet的heatmap会不会更好？？？
      • focal loss $L_{loc}$
    • anchor shape prediction
      • predicts the best shape for each location
      • best shape：a shape that lead to best iou with the nearest gt box
      • 一层卷积：1x1 conv，channel2，[-1,1]
      • transform layer：transform direct [-1,1] outputs to real box shape
        • $w = \sigma s e^{dw}$
        • $h = \sigma s e^{dh}$
        • s：stride
        • $\sigma$：经验参数，8 in experiments
      • set 9 pairs of (w,h) as RetinaNet，calculate the IoU of these sampled anchors with gt，take the max as target value
      • bounded iou loss：$L_{shape} = L_1(1-min(\frac{w}{w_g}, \frac{w_g}{w})) + L_1(1-min(\frac{h}{h_g}, \frac{h_g}{h}))$
    • feature adaptation
      • intuition：the feature corresponding to different size of anchor shapes应该encode different content region
      • inputs：feature map & anchor shape
      • location-dependent transformation：3x3 deformable conv
      • deformable conv的offset是anchor shape得到的
      • outputs：adapted features
    • with adapted features
      • then perform further classification and bounding-box regression
    • training
      • jointly optimize：$L = \lambda_1 L_{loc} + \lambda_2 L_{shape} + L_{cls} + L_{reg}$
      • $\lambda_1=0.2，\lambda_2=0.5$
      • each level of feature map should only target objects of a specific scale range：但是ASFF论文主张说这种arrange by scale的模式会引入前背景inconsistency？？
    • High-quality Proposals
      • set a higher positive/negative threshold
      • use fewer samples