简体繁体 English

如何去除 Faster RCNN 中的误检（误报）

[英]How to remove false detection (False Positives) in Faster RCNN

原文 2021-04-08 02:29:17 1 1 performance/ tensorflow/ object-detection/ faster-rcnn

I am using Faster RCNN with Inception V2 on custom dataset.我在自定义数据集上使用 Faster RCNN 和 Inception V2。 My model is working fine with good detection accuracy.我的 model 工作正常，检测精度高。 However, I am facing false positive problem when I pass an image to the model I get correct prediction but I am also getting some wrong bounding boxes with high confidence score.但是，当我将图像传递给 model 时，我遇到了误报问题，我得到了正确的预测，但我也得到了一些错误的高置信度的边界框。 Is there any method which can be used as a post-processing to remove these extra detection?是否有任何方法可以用作后处理来消除这些额外的检测？

1 个解决方案

It seems like this is a common problem with Transfer Learning, you should check out this discussion.这似乎是迁移学习的一个常见问题，你应该看看这个讨论。 In the end it seems that all boils down to what's the source of your false positives.最后，似乎一切都归结为您误报的来源。

For instance, once I trained a detector to detect smoke in wildfire images, but it ended up also catching the clouds.例如，有一次我训练了一个检测器来检测野火图像中的烟雾，但它最终也捕捉到了云层。 To solve that I also annotated the clouds as a new class, and ignored it's detections.为了解决这个问题，我还将云注释为新的 class，并忽略了它的检测。 This greatly improved the performance.这大大提高了性能。

If it's making wrong detections with high confidence I think it would be hard to solve this problem with only a post-processing.如果它以高置信度进行错误检测，我认为仅通过后处理很难解决这个问题。

You could also try hard mining.你也可以尝试努力挖矿。 Altough I'm not really sure how to do it for Faster R-CNN.虽然我不太确定如何为 Faster R-CNN 做到这一点。