EMDLoss：取min问题 about crowddet HOT 8 OPEN

As described in the paper, A proposal generally need to infer 2 different detections, and there are 2 different combination choices among the detections and their targets, namely det_loss((B_0, G_0)) + det_loss((B_1, G_1))), det_loss((B_0, G_1)) + det_loss((B_1, G_0)), then a minimum operation performs on those losses to select the minimum one for updating parameters. But we hadn't noticed your hypothesis that a detection head may produce a loss largely higher than the other one. It can be attributed to the reason that a normalizer will scale the loss what is identical to the general detection loss functions.

However, what we have observed is that sometimes the first head will produce the majority of true positives while the second head infers the supplemental true positives, and sometimes the second one will predict the majority of true positives, and the first one supplements the rest of true positives. For the reason that the number of ground-truth for training the first head and the second head is different during the label assignment strategy, but the EMD loss has the property of permutation-invariance.

from crowddet.

As described in the paper, A proposal generally need to infer 2 different detections, and there are 2 different combination choices among the detections and their targets, namely det_loss((B_0, G_0)) + det_loss((B_1, G_1))), det_loss((B_0, G_1)) + det_loss((B_1, G_0)), then a minimum operation performs on those losses to select the minimum one for updating parameters. But we hadn't noticed your hypothesis that a detection head may produce a loss largely higher than the other one. It can be attributed to the reason that a normalizer will scale the loss what is identical to the general detection loss functions.

However, what we have observed is that sometimes the first head will produce the majority of true positives while the second head infers the supplemental true positives, and sometimes the second one will predict the majority of true positives, and the first one supplements the rest of true positives. For the reason that the number of ground-truth for training the first head and the second head is different during the label assignment strategy, but the EMD loss has the property of permutation-invariance.

所以关键部分就是如何分配G_0和G_1我看您是通过iou评分来划分的, G_0为iou评分最高的gt框而G_1为iou次高的gt框, 但是在yolov5系列算法中, gt框是直接通过wh的比值来确定去留的, 直观的分出两类gt。我还有一个疑问， 直接通过iou来划分G_0和G_1的话是否也会存在同一个label同时被两个头预测的情况呢?我想此系统在Deepsort上工作, 重叠的检测框会带给我困扰

from crowddet.

在同类遮挡的情况下,两个检测头会预测相同的label,这种情况下2个检测头预测的label虽然相同,但是检测框分别对应2个不同的同类物体.如果不存在遮挡情况,那么只会有一个检测头预测一个label,剩下的一个检测头对应的label是背景;

from crowddet.

在同类遮挡的情况下,两个检测头会预测相同的label,这种情况下2个检测头预测的label虽然相同,但是检测框分别对应2个不同的同类物体.如果不存在遮挡情况,那么只会有一个检测头预测一个label,剩下的一个检测头对应的label是背景;

但是按照setNMS的机制如果同一层的一个框被选中, 另一个背景框岂不是也会被强制复活?

from crowddet.

所以在进行set NMS的时候会卡一个置信度阈值0.05,在一般情况下另一个背景框会被移除.如果背景框在卡阈值阶段无法被移除,最后也只能被强行复活成为FP.所以我们设计了refinement module.但这种情况一般很少出现.

from crowddet.

所以在进行set NMS的时候会卡一个置信度阈值0.05,在一般情况下另一个背景框会被移除.如果背景框在卡阈值阶段无法被移除,最后也只能被强行复活成为FP.所以我们设计了refinement module.但这种情况一般很少出现.

了解了，十分感谢您解答我的疑问，谢谢

from crowddet.

如果一个proposal的label assignment按照这种分配方式的话，会不会存在一种潜在的问题呢？就是proposal的第一部分的特征和输出结果其实与GT2相关度比较高，但是分给他的label是GT1，从而影响了模型训练

from crowddet.

from crowddet.