一种用于目标跟踪边界框回归的光滑IoU损失

Smooth-IoU Loss for Bounding Box Regression in Visual Tracking

边界框回归分支是深度目标跟踪器的关键模块,其性能直接影响跟踪器的精度.评价精度的指标之一是交并比(Intersection over union,IoU).基于IoU的损失函数取代了l_(n)-norm损失成为目前主流的边界框回归损失函数,然而IoU损失函数存在2个固有缺陷:1)当预测框与真值框不相交时IoU为常量0,无法梯度下降更新边界框的参数;2)在IoU取得最优值时其梯度不存在,边界框很难收敛到IoU最优处.揭示了在回归过程中IoU最优的边界框各参数之间蕴含的定量关系,指出在边界框中心处于特定位置时存在多种尺寸不同的边界框使IoU损失最优的情况,这增加了边界框尺寸回归的不确定性.从优化两个统计分布之间散度的视角看待边界框回归问题,提出了光滑IoU(Smooth-IoU,SIoU)损失,即构造了在全局上光滑(即连续可微)且极值唯一的损失函数,该损失函数自然蕴含边界框各参数之间特定的最优关系,其唯一取极值的边界框可使IoU达到最优.光滑性确保了在全局上梯度存在使得边界框更容易回归到极值处,而极值唯一确保了在全局上可梯度下降更新参数,从而避开了IoU损失的固有缺陷.提出的光滑损失可以很容易取代IoU损失集成到现有的深度目标跟踪器上训练边界框回归,在LaSOT、GOT-10k、TrackingNet、OTB2015和VOT2018测试基准上所取得的结果,验证了光滑IoU损失的易用性和有效性.

The branch of bounding box regression is a critical module in visual object trackers,and its performance directly affects accuracy of a tracker.One of evaluation metrics used to measure accuracy is intersection over union(IoU).The IoU loss which was proposed to replace l_(n)-norm loss for bounding box regression is increasingly popular.However,there are two inherent issues in IoU loss:One is that the parameters of bounding box can not be updated via gradient descent if the predicted box does not intersect with ground-truth box;the other is the gradient of the optimal IoU does not exist,so it is difficult to make the predicted box regressed to the IoU optimum.We reveal the explicit relationship among the parameters of IoU optimal bounding box in regression process,and point out that the size of a predicted box which makes IoU loss optimal is not unique when its center is in specific areas,increasing the uncertainty of bounding box regression.From the perspective of optimizing divergence between two distributions,we propose a smooth-IoU(SIoU)loss,which is a globally smooth(continuously differentiable)loss function with unique extremum.The smooth-IoU loss naturally implicates a specific optimal relationship among the parameters of bounding box,and its gradient over the global domain exists,making it easier to regress the predicted box to the extremal bounding box,and the unique extremum ensures that the parameters can be updated via gradient descent.In addition,the proposed smooth-IoU loss can be easily incorporated into existing trackers by replacing the IoU-based loss to train bounding box regression.Extensive experiments on visual tracking benchmarks including LaSOT,GOT-10k,TrackingNet,OTB2015,and VOT2018 demonstrate that smooth-IoU loss achieves state-of-the-art performance,confirming its effectiveness and efficiency.