Tensorflow对象检测API Faster-RCNN收敛但检测不准确

Question

I am trying to use the Tensorflow Object detection API to recognize the Guinness logo. The process is similar to that shown here - https://towardsdatascience.com/building-a-toy-detector-with-tensorflow-object-detection-api-63c0fdf2ac95 .

I have prepared 100 training images, which I use augmentation to reach a total of around 5000 training images. (Using Imgaug). In tensorboard when training I see what looks like a good learning curve, reaching a loss of < 0.1 , but when I export and test the graph I get lots of false positives and very inaccurate results. I am trying to work out why this is.

Tensorboard performance graphs

Bad detection example

Note, to automate labelling of my images I cropped the original 100 neatly around the logo, then I programmatically place them on a random background image, with the bounding box around it. Example -

Like so -

All the training images are 800x600, but the actual bounding box and logo would be much smaller as you can see.

And here is the xml annotation file for that same image -

 <?xml version="1.0" encoding="utf-8"?> <annotation> <folder>images</folder> <filename>57.png</filename> <path>model\\images\\57.png</path> <source> <database>Unknown</database> </source> <size> <width>800</width> <height>600</height> <depth>3</depth> </size> <segmented>0</segmented> <object> <name>guinness</name> <pose>Unspecified</pose> <truncated>0</truncated> <difficult>0</difficult> <bndbox> <xmin>225</xmin> <ymin>329</ymin> <xmax>516</xmax> <ymax>466</ymax> </bndbox> </object> </annotation> 

Does anybody know why tensorflow would be correctly classifying the test images, but also have inaccurate detection when I test on a real world image? Any advice is welcome, and feel free to ask for more information.

Answer 1

Couple of thoughts

• Is your test image also of same size 800*600?

• You probably want to play with the image_resizer value in the config file

Answer 2

Eventually I gave up the method of placing my logo images onto a random background, instead I manually labelled them, and then used image augmentation to increase my training set size. This seemed to greatly improve my results. I think this has something to do with a contextually accurate background actually being quite important in training.

Hopefully this is helpful to some, thanks for the help.