ICCV-2017
NMS 是许多目标检测算法重要的步骤之一
NMS 的缺点,if an object lies within the predefined overlap threshold, it
leads to a miss.(sets the score for neighboring detections to zero)
Intuitively, if a bounding box has a very high overlap with M M M(maximum score), it should be assigned a very low score, while if it has a low overlap, it can maintain its original detection score.
作者对 NMS 进行了改进
NMS first employed in edge detection techniques
NMS 与 maximum score box 的 IoU 大于阈值的都抑制为了 0
全抑制为0,容易漏检,NMS should take the following conditions into account,
高 NMS 阈值情况下,mAP 反而会下降,因为 the increase in false positives would be much higher than the increase in true positives for this case because the number of objects is typically much smaller than the number of RoIs generated by a detector.
soft-NMS 大于阈值的分数降低
a list of detection boxes B B B with scores S S S.
maximum score M M M
the set of final detections D D D
suppressing all nearby detection boxes with a low N t N_t Nt would increase the miss-rate.
其中 Rescoring Functions f ( i o u ( M , b i ) ) f(iou(M, b_i)) f(iou(M,bi)) 的形式有如下两种
(1)非连续的(not continuous)
与最高分box 的 IoU 越大,分数降得更多(误检的概率比较大),IoU 越小,分数降的越少
IoU 等于 N t N_t Nt 的时候,值不一样
also a greedy algorithm and does not find the globally optimal re-scoring of detection boxes.
不一定局限于指数形式,也可以是 Gompertz function
N t N_t Nt to 0.3, σ \sigma σ to 0.5
Soft-NMS has more potential to improve recall at higher O t O_t Ot(detection evaluation threshold).
试试 soft-NMS 超参数 σ \sigma σ 的敏感性
(1)Localization Performance
千言万语,不如下面一个表来的直接
even when we obtain better performance at higher O t O_t Ot, performance at lower O t O_t Ot does not drop.
low values of σ \sigma σ perform better at lower O t O_t Ot and higher values of sigma perform better at higher O t O_t Ot.
(2)Precision vs Recall
a large wide bounding box spanning multiple people is suppressed because it had a small overlap with multiple detection boxes with a higher score than it.
soft-NMS也是一种贪心算法,并不能保证找到全局最优的检测框分数重置。