Faster R-CNN training correctly but the detector finding bizarre regions during testing

1 次查看（过去 30 天）
Josh
Josh 2019-3-12
编辑： Josh 2019-3-12
Hi all,
I'm having some troubles with the detect function when using matlab's implementation of Faster R-CNN for deep learning.
I'm training on 300+ images with somewhere between 50-160 regions labeled in each image for ground truth. Using vgg16 as the base network and replicating parameters from the same model which is working in a different language (I have since played around with these parameters a bit). The training appears successful, minibatch accuracy is good, there's no erors and I get a detector object which I can save as a .mat file of approximately 400 MB.
The problem is that as soon as I apply the detector to an image from the test set, or even the same images from the training set, every single box detected has the same score, and they're tiled uniformly throughout the image, left to right, top to bottom, up until a bout a third of the way through the image where they just stop. Perfectly uniform, no regard for what's in the test image at all.
I've tried the training and testing in 2018b and 2019a for the same results. I've tried tuning the parameters in a number of ways, resizing the original images and so on. The result is always the same.
Has anyone else had this type of issue with the Faster R-CNN functionality, or perhaps any ideas on what might be causing it?
Thank you in advance!

请先登录，再回答此问题。

回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Recognition, Object Detection, and Semantic Segmentation 的更多信息

标签

产品


版本

R2018b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by