CNN Object detection: How to reduce high false positive rates

Question

I am currently using Faster RCNN with inception v2 on Tensorflow object detection API on WIDER-FACE data, i have a lot of false positives with high scores (>0.98, so setting a higher score threshold won't help). I have already assigned Hard Example Mining in my code but it doesn't help to much. For image preprocessing, I randomly crop an area from the original image using tf.image.sample_distorted_bounding_box and resize it to 300*300. The resized image will be randomly flipped by a probability of 0.5

I set batch size to 32 and for each image the ratio of positive:negative in hard negative mining operation is 32:32 . I set IoU > 0.5 as positive and IoU < 0.3 as negative. For samples that IoU between 0.5 and 0.3 are ignored

Can anyone help me with this? Many thanks!

Answer 1

Usually the ratio between positive and negative examples is set to be 1:3. It could be that because in your model the ratio is set to 1:1, the model doesn't see enough negative examples.

You can also try to do an error analysis and check what are the false positives which have high confidence. Maybe they are of a specific type, and if so - add negative examples of this type to your training data.

Answer 2

您也可以尝试增加 RPN 的 NMS 阈值。