Then the rest is up to us! Our mobile technician will arrive on time and completely prepared to get your car up and running again. Go ahead and book a mobile wheel bearing replacement with Wrench today!

box regression output是一个WxHx4大小的feature map，那么在坐标为（i，j）的点上是一个长度为4的向量，分别表示4个偏移量。假设一个instance，其在feature level为 l l l对应的有效区域为 b e l b{^l_e} bel​，这个instance会影响到 b e l b{^l_e} bel​内的每个值。假设坐标为（i，j），长度为4的向量 x i , j l x^l_{i,j} xi,jl​表示这个instance的上、左、下、右4边界和（i，j）的距离，然后 x i , j l / S x^l_{i,j}/S xi,jl​/S作为最后输出结果，S是一个标准化常量，我们选择S=4。分支采用IoU Loss，整个regression loss是一张图片中每个有效区域的IoU Loss的均值。

参考链接： 知乎: CVPR2019 | 目标检测 FSAF：为金字塔网络的每一层带去最好的样本. 知乎: [CVPR2019]:FSAF for Single-Shot Object Detection. 部分翻译: 论文阅读笔记四十六：Feature Selective Anchor-Free Module for Single-Shot Object Detection（CVPR2019）.

Although wheel bearings are engineered to last the entire lifespan of your vehicle, they can break and get worn out. Most wheel bearings last about 85,000 to 100,000 miles before they need to be replaced. Some vehicles have wheel bearings that last a lot longer, and some vehicles need them replaced fairly quickly. It all depends on the quality of your wheel bearings and your usual driving conditions.

A wheel bearing is a set of steel balls or tapers that we call "ball bearings" or "tapered bearings." These bearings are held together by a metal ring called a "race." In the center of the wheel, there is a hollow piece of metal called a "hub," and the wheel bearings fit inside the hub to help rotate your wheels.

If you ignore bad wheel bearings, your wheel could stop working while driving, or you could have issues steering your car.

The first thing we will need to know is what kind of vehicle you have. You should provide us with the year, make, and model, as well as the VIN and license plate number. This information will help us properly prepare for your wheel bearing replacement and any other services you may have requested. If your car requires a specific part, we will make sure we come with the correct tools and parts to perform all of your necessary repairs and services.

为解决目标检测的一大挑战scale variation，通常可用特征金字塔网络 FPN（Feature Pyramid Network）。但它有一个问题：不同尺寸的物体依据其与 FPN 每一层 Anchor 的适配程度，分配到不同分辨率的层上进行学习，期望其能够充分适配各层感受野和空间信息，从而使得检测器能够检测尺寸不一物体；这种启发式的方法（heuristic-guided feature selection）虽听着合理，却没有明确的证据，我们并不知道一个确切的 ground-truth 分配到哪一层上去学习是最合适的。如图2，一个大小为50×50像素的car实例和另一个大小为60×60像素的类似car实例可以被分配到两个不同的特征层，而另一个大小为40×40的car实例可以被分配到与50×50实例相同的层。 因此，要想取代这种分配方式，就必须不再依赖每一层 anchor 和 ground-truth 的 IoU 来做分配。因此本文提出的模块也是就叫做 Feature Selective Anchor-Free Module（FSAF），让网络自己学习该怎么分配，为金字塔网络的每一层带去最好的样本。 怎么做？ 在此之前，我们先温习一下具有 FPN 式结构的检测器中是如何分配 ground-truth 的。我们把 FPN 上所有层的所有 anchor 全部拿过来，和 ground-truth 计算 IoU overlap，通过设定的阈值或者是选择最大的 IoU，便可以得到这个 ground-truth 的 positive anchor。由于每一层上的 anchor 都是密集设置的，层与层之间仅仅是 anchor 的大小不同（随着分辨率的升高，anchor 的尺寸下降），因此这个分配过程基本就是看 ground-truth 的大小：小的 ground-truth 被分配到高分辨率，大的 ground-truth 被分配到低分辨率。这样的设计作者认为它是 heuristic（启发式） 的，可能会使得选择的特征不是最优的，从而影响检测器的性能。

论文地址: https://arxiv.org/abs/1903.00621. 作者在B站解读视频：https://www.bilibili.com/video/av49972561/. 注：本文在查阅时发现了两篇知乎专栏写的很好的解读，在文末给出链接，因此这里就仅仅给出整理，并添加了一些细节性的东西。 先看看效果，论文提出的特征选择无锚点模块(FSAF)应用在单阶段检测器RetinaNet上，使用ResNeXt-101在COCO数据集上达到单阶段SOTA，map提升了1.8%仅多用了6ms。如图，FSAF对小的物体效果更好。

Then we will need to know what auto repair or maintenance service you need. If you know your wheel bearings have gone bad, book a wheel bearing replacement appointment! If you suspect that your wheel bearings are the cause of your automotive issues, but you aren't sure, book a diagnostic appointment and our mobile technician will take a look at your car and figure out what's going on.

There are many signs you should look out for if you are concerned about your wheel bearings. If you notice any of the signs listed below, you might need a wheel bearing replacement. Keep your eyes and ears alert if you suspect your wheel bearings are going bad, and let your mechanic know which signs you've noticed.

We are back with another market announcement! If you have been following Wrench from the start, you will know that 2021 is a year of growth for us. We are opening new markets

这里插一段先介绍一下图4.它展示了带有FSAF模块的RetinaNet。总的来说，RetinaNet由一个骨干网（图中未展示）和两个子网络（分类和回归）组成。特征金字塔由骨干网构建，从 P 3 P_3 P3​到 P 7 P_7 P7​， l l l是金字塔层数， P l P_l Pl​的分辨率是输入图像的 1 / 2 l 1/2^l 1/2l。在每层 P l P_l Pl​的后面添加分类及回归子网络，二者都为全卷积网络。 anchor-based分支：分类网络预测每个位置的A个anchor及K个类别的分数值，回归分支预测距离存在的最近的A个anchor框的四维未知类别的offset。 anchor-free分支： 一个3x3xK的卷积层添加到分类分支的输出，后接sigmoid函数，与基于anchor分支的部分相互平行，其预测目标物在每个位置上K个类别的概率值。一个3x3x4的卷积层添加到回归分支的输出，其后接ReLU函数。其作用是以anchor-free的方式预测框的偏移量。 anchor-based与anchor-free共享每层特征以多任务的方式进行运作。 所以，FSAF仅仅就是！！！！！！！ 在 RetinaNet 的 box 和 cls 分支上仅仅各加了一层 conv layer，分别生成一个 W × H × K classification output 和一个 W × H × 4 的 regression output。 了解了基本网络框架，所以它具体是怎么对应GT计算损失的呢？

To keep our customers educated about their car and its various parts, we will go over everything you need to know about wheel bearings and why they are important for your vehicle. Then you will be able to quickly identify wheel bearing problems in the future, or at least know when it is time to visit a mechanic.

In addition to the danger, it can put a lot of strain on other parts of your vehicle, such as the hub, CV joint, and transmission. If you don't repair your failing wheel bearings soon enough, you might have to spend even more money on auto repair services. You might also notice that your tires wear out faster, which will mean you will need to purchase new tires more often. Overall, all of these potential side effects of bad wheel bearings can end up being more expensive than the original problem.

（1）对于一个instance，ground truth box的类别： k k k； （2）ground truth bounding box坐标： b = [ x , y , w , h ] b=[x,y,w,h] b=[x,y,w,h]，其中， ( x , y ) (x,y) (x,y)表示box的center坐标； （3）ground truth bounding 在第 l l l个feature level上的投影坐标： b p l = [ x p l , y p l , w p l , h p l ] b{^l_p}=[ x{^l_p}, y{^l_p}, w{^l_p}, h{^l_p}] bpl​=[xpl​,ypl​,wpl​,hpl​]； （4）effective obx: b e l = [ x e l , y e l , w e l , h e l ] b{^l_e}=[ x{^l_e}, y{^l_e}, w{^l_e}, h{^l_e}] bel​=[xel​,yel​,wel​,hel​]，它也表示 b p l b{^l_p} bpl​的一部分，缩放比例系数为 ϵ e = 0.2 \epsilon_e=0.2 ϵe​=0.2（图5中的白色部分）； （5）ignoring obx: b i l = [ x i l , y i l , w i l , h i l ] b{^l_i}=[ x{^l_i}, y{^l_i}, w{^l_i}, h{^l_i}] bil​=[xil​,yil​,wil​,hil​]，它也表示 b p l b{^l_p} bpl​的一部分，缩放比例系数为 ϵ i = 0.5 \epsilon_i=0.5 ϵi​=0.5（图5中的灰色部分）. classification output是一个WxHxK大小的feature map，K表示物体类别数，那么在坐标为（i，j）的点上是一个长度为K的向量，表示属于每个类别的概率。分支对应的gt是图中白色区域内值为1，表示正样本，黑色区域内值为0，表示负样本，灰色区域是忽略区域不回传梯度。分支采用Focal Loss，整个classification loss是非忽略区域的focal loss之和，然后除以有效区域内像素个数之和来正则化一下。

Luckily for you, a wheel bearing replacement service isn't a super complicated job, especially for professional technicians. This means it won't be as expensive or time-intensive as a more severe issue. However, if you ignore this issue for too long, other parts might be affected by the bad wheel bearings, adding more repairs and costs to your final bill.

在inference阶段，对classification output和regression output的输出进行decode。在每个像素位置 （ i ， j ） （i，j） （i，j），假定预测坐标是 [ O t i , j − , O l i , j − , O b i , j − , O r i , j − ] [O_{t_i,j}^-,O_{l_i,j}^-,O_{b_i,j}^-,O_{r_i,j}^-] [Oti​,j−​,Oli​,j−​,Obi​,j−​,Ori​,j−​]，预测距离是 [ S O t i , j − , S O l i , j − , S O b i , j − , S O r i , j − ] [SO_{t_i,j}^-,SO_{l_i,j}^-,SO_{b_i,j}^-,SO_{r_i,j}^-] [SOti​,j−​,SOli​,j−​,SObi​,j−​,SOri​,j−​]，预测映射框是 [ i − S O t i , j − , i − i − S O l i , j − , i − S O b i , j − , i − S O r i , j − ] [i-SO_{t_i,j}^-,i-i-SO_{l_i,j}^-,i-SO_{b_i,j}^-,i-SO_{r_i,j}^-] [i−SOti​,j−​,i−i−SOli​,j−​,i−SObi​,j−​,i−SOri​,j−​]，最后乘以stride = 2 l 2^l 2l就得到了在image上的位置坐标。框的置信度得分和类别可以由classification output maps上位置 （ i ， j ） （i，j） （i，j）处的k维向量的最大得分和对应的类别来决定。

Your car's wheels are very important as you won't be able to drive anywhere without them. Unless you are a mechanic or you are an avid car enthusiast yourself, you might not know how many parts go into a wheel. There is a tire, rotor, bearing, stator, and capacitor ring, just to name a few. The average car owner doesn't even know the purpose of parts such as wheel bearings, let alone know when they are failing!

Looking for a mobile mechanic to replace your wheel bearings? Look no further because Wrench is here to help! We have hundreds of ASE-certified mobile technicians all over the country that have the skills to replace your wheel bearings or perform any requested auto repair or maintenance service.

Your car's exhaust system, and all of the various parts, including the exhaust manifold, are important to keeping your car on the road. Without a properly functioning exhaust system, you could

No. We don't recommend driving without a properly functioning wheel bearing. Not only is it extremely dangerous for you, but you can also end up injuring or killing other drivers on the road.

很明显，要想取代上述的分配方式，那么“把 FPN 上所有层的所有 anchor 全部拿过来，和 ground-truth 计算 IoU”这一步就不能有。因此作者提出了一个 anchor-free 的 module，称为 FSAF，大体思想是在每一层都插入这个模块，尝试不用 anchor 去检测 instance，而后看看哪一层的 FSAF 对于这个 instance 的损失最小，不就可以认为这一层是最适合检测这个 instance 的吗？ 接着 FSAF 大喊一句“安排”！！！把这个物体安排到这个层，再用 anchor-based 的模块去检测.我们将提出的FSAF模块嵌入到retinaNet，实际上它可嵌入到任何带有特征金字塔结构的的单阶段检测器。

Log into our mobile app or website to book a wheel bearing replacement service today! You can book an appointment as a guest if you don't have an account with us just yet. Whether you have an account with us or not, we will need to gather a little information about your type of vehicle and its needs.

在training中，采用multi-task loss，即 L = L a b + λ ( L c l s a f + L r e g a f ) L=L^{ab} +\lambda(L^{af}_{cls} + L^{af}_{reg}) L=Lab+λ(Lclsaf​+Lregaf​)，其中 L a b L^{ab} Lab是原本anchor-based RetinaNet的总损失， L c l s a f L^{af}_{cls} Lclsaf​和 L r e g a f L^{af}_{reg} Lregaf​分别是anchor-free分支的总分类和回归损失， λ \lambda λ控制anchor-free分支的权重，我们设置 λ = 0.5 \lambda=0.5 λ=0.5.在推理(inference)时不需要通过 λ \lambda λ做选择。

Since your wheel bearings are located on your wheel, they need to be able to withstand harsh driving and weather conditions such as potholes, speed bumps, rain, dust, and more. All these harsh conditions can reduce their lifespan, so you should keep an eye out for the symptoms of a failing wheel bearing rather than relying on a timeline.

Once we know which services your car needs and what kind of car you have, we will calculate an auto repair quote for you. This quote will be personalized just for you, so you can know exactly what to expect. We will include everything from the cost of the parts and labor to the estimated time it will take to complete the services. After you have had the chance to look over our auto repair quote, you will pick a date, time, and location for your wheel bearing replacement appointment.

Some of these symptoms of bad wheel bearings are very serious and need immediate attention. Before hopping back on the road, make sure to get your car checked out by a qualified mechanic!

ablation study部分分析了anchor-free的必要性，online feature selection的重要性，以及选择的feature level是不是最优的。同时还指出FSAF非常robust和efficient，在多种backbone条件下，都有稳定的涨点。在ResNext-101中，FSAF超过anchor-based1.2个点，同时快了68ms，在AB+FSAF情况下，超过RetinaNet1.8个点，只慢了6ms，效果也是非常显著的。具体参见论文。