I am training a SSD model for detecting mobile cranes. The training dataset contains 1,000 images and test set over 400 images. About 200 epochs gave mAP 83%, but my target is 90%. So I trained SSD-ResNet-101 and it gave less accuracy.

I assume that it is because ResNet-101 is too deep for the size of my dataset. I consider using ResNet-50 and Inception. But I don't have time to experiment all the models with different parameter settings.

Is there anyone who has experience in this direction? Any advice is welcome.

Thanks in advance.

  • $\begingroup$ Yes as you said the dataset is likely to be too small for that amount of layers and the model is likely to overfit. I would try with shallower architectures. Inception and ResNet are better for fine tuning in your case. $\endgroup$ Oct 3 '18 at 12:47
  • $\begingroup$ Thank you, @FrancescoPegoraro. Do you mean ResNet-50? $\endgroup$
    – IPRNG
    Oct 3 '18 at 14:46

In general, the best way to increase performance to increase training data and to train longer.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.