0

so i'm trying to make object detection model based on Keras RetinaNet and ResNet-152 backbone. I have followed every tutorial that explain how to do it. Up untill now, i have trained my model where the loss i achieve is 0.81, i thought it was an excellent result.

Classification Lost

Regression Lost

Here's the example of bounding box process i did in label img

But somehow, when i tried to predict the result to new images (4 images), the result is really weird, here's the each of image result

Filename, boxes, scores, labels

Did i do something wrong in dataset creation? because i only tweak some code, like

  • i comment "setup_gpu(gpu)" command
  • i change the default backbone which is from "resnet50" to "resnet152" in train.py argument section

For the code, i used https://github.com/fizyr/keras-retinanet

Edit :

These are the examples when i tried to predict even with my own training data. Althought one of them is correct, but the confidence score is very low, it's not even reached 0.5.

  1. Prediction - I
  2. Prediction - II
Mahesa A. N E. S
  • 1
  • 1
  • I don't think it makes any sense to look at the raw predictions like that, you should visualize bounding boxes with some code (after thresholding predictions and applying NMS). – Dr. Snoopy Apr 09 '21 at 23:49
  • Hi Dr Snoopy, thanks for replying, i have edited my questions and attached instance of predicted images, do you have any idea why it's not working? thanks in advance. – Mahesa A. N E. S Apr 10 '21 at 00:40
  • Why do you think the predictions are not as expected? Your prediction files indicate that some bounding boxes were identiifed. It would be better if you could plot those predicted boxes on the original image to get a better visualization. – rohit_r Apr 12 '21 at 11:47

0 Answers0