Training MXNet Faster R-CNN on own data set

Question

I was using the mxnet example for Faster R-CNN on the official GitHub: Faster R-CNN

I created my own dataset and adapted the pascal_voc.py file. This included changing the number of classes to 13. The resolution of my pictures is 600*800 pixel, so a bit larger than the VOC dataset. In my trainval images, I have 2000 examples, so 3000 less than than in VOC. I am using mxnet 0.10 in python for this.

I am getting no error in the training, the loss is decreasing, but when I use the model after 10 epochs, I am getting no results, just the:

class ---- [[x1, x2, y1, y2, confidence]]

Does anyone has an idea what I am might missing?

What do you mean "use the model...", what code are you running? You will get this output if the model didn't find any box in the input. — Guy, Jul 04 '17 at 16:07
Hi Guy, I am running the example from github that I mentioned in the link. But instead of the 21 original classes, I defined 13 classes (Faces). Yes that I know, and I am now very sure that I am lacking training data (My dataset is too small) — Max, Jul 05 '17 at 09:27

score 0 · Answer 1 · answered Sep 11 '18 at 18:34

0

If you are comfortable with gluon, a flexible interface of mxnet, do you mind to try the Faster RCNN example from GluonCV:

https://gluon-cv.mxnet.io/build/examples_detection/train_faster_rcnn_voc.html

answered Sep 11 '18 at 18:34

TomHall

286
3
15

Training MXNet Faster R-CNN on own data set

1 Answers1