I trained an object detection model with 5K images, it works most of the time, but I am facing an issue, for few times, the object is not getting detected.

So, I planned to retrain the model, for that I collected around 1K images where the trained-model is not detected the objects in these 1K images, and annotated using rectangle bounding box.

Now, my clarification is - should I add the newly collected images in the validation dataset or in training dataset.

To experiment it, instead of splitting the images for training and validation, I planned to add all the images validation dataset because the quantity of the objects in each category is less around 100 to 200. Totally 11 classes of objects are there.

My thought is that, if I add the new images in the validation dataset, then it would be better to improve the model, because this 2nd time training is for fine tuning and also the model is updating the weights based on the validation dataset, so, I am thinking, the model would be improved much.

And also, I am had an discussion with my friend, they mentioned to add the new images in the training dataset so that it would be improved instead of adding in the validation dataset

I am in confusion state, could you please clarify this - in which dataset[training or validation dataset] shall I add all the new images so that the model will improve


  • $\begingroup$ split 70/30 train/validation. But keep in mind that retraining for specific object might result in reducing performance for other objects. $\endgroup$
    – Nikos M.
    Mar 15 at 6:13


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.