Do you guys have experience with object detection models?

Can I validate my object detection model with a classifier or another model?
Object detection model may do false detects sometimes, but it is fast.
Can I use another model which is slow but accurate and validate if the detection was correct or not?

1 Like

Detecting distance or depth generally requires either:

  • two images from offset viewpoints (consider why you have two eyes that are a few centimeters apart),
  • or a set of images from one camera where the focus is set to different distances.

I meant I want to detect it at every distance, I am not trying to measure its distance.
I wonder if I can validate my detections with another model or something.
Thanks for your answer, sorry for misunderstanding.

It is difficult to identify the same object at different distances (which appear to be a smaller size with lower resolution), unless your training set has examples of the same object at different sizes.

I don’t know exactly what you mean by “validate my detections with another model”.

If you distrust the ‘fast’ model so much that you have to run a second model, why use the first one at all? Maybe I am missing something, but it seems like there is no benefit to the first model whatsoever, only cost. Is throughput important in your business problem? If not, just run the slow model you trust. If so, the slow but accurate doesn’t help. Instead, you need to look at what is causing the poor accuracy of the fast model and find ways to improve it. Look closely at what parts of the confusion matrix are not satisfactory. Is it ok in training but not during operation? Does it also underperform during training? What specific types of objects does it improperly detect? These analyses might suggest an approach to improved detection performance.

I detect an object and send it to tracker, I have to track it for a while to say that I am succesful. But If I detected wrong object and I track it for a while, like 5 seconds, I am unsuccesful.
So, I need to detect an object when it is available as fast as possible and start tracking it.
But also I should stop tracking it if it is not the object that I should track asap.

I’m not sure if it is technically possible or if it makes sense, but I can’t find any other solution yet.

One fast model which detects object when it can do, one accurate model which stops tracking if its wrong object.
Because object doesn’t stay in the screen so long and if I don’t use fast one, I thought it may be too late to detect it.

This models works real time btw.

I can’t build a fast and also accurate model yet so I thought two models, one fast and another one accurate could work together :smile:

In my experience, if you are using a strong object detector, training a classifier on the top with all the classes does not help much (assuming the model was well trained), but it depends on the model size and the problem. If you train a huge model with a lot of data or you have a specific problem it may help. However, what I have found is that most of the false positives I got were objects that resembled those used to train the model. So, what has helped me is to add binary classifiers to those classes that had the most false positives.

But you should try different types of object detectors and also tweak the parameters and augmentations in case the problem can be solved directly.

Hope that helps you.


I am detecting just one kind of object, nice to hear that binary classifiers helps to decrease false positives! Tysm

Hello @mrtckaya

Based on my understanding if you have just one kind of object with focus is on feature extracting and information is sparse then introduce additional layers in the ‘classifier_layers’ function like adding Global Average Pooling layer which would spatially summarizing the features, followed by Fully Connected layers to increase the model’s learning capacity.


I’ll do a research about what you said, thank you so much!