Object Detection

Hello. I just want to find the possible positions of objects. When I tried this with edge detection methods, they performed very badly (in terms of detection) in very noisy areas. I also used models like HED and they also had performance (speed) problems. I didn’t want to do something like this as it is necessary to train models like YOLO. Does anyone have any advice? Thank you very much.

1 Like

may provide some processed pictures?

1 Like

Of course:

how can i detect possible positions of objects?

and since processed images already are unsalvageable, please also provide the unfiltered source

unfiltered image. i want to detect objects in this photo like keyboard etc.

so you should use a neural network for object detection.

you don’t need to train anything. why do you think you need to train anything?

but neural networks can only detect certain objects. are there models that will detect every possible object? i have not come across any except graund dino. graund dino is an extremely large model and i don’t need that much size.

are you implying that anything less than a neural network will somehow be able to successfully distinguish more classes? that makes no sense.

there are trained models with large sets of classes.

besides that, you can always take a trained model and retrain the classification layer (last layer) to your specific object types.

there is no magic bullet. you either know/learn this stuff, or you have to spend compute time on a very large model with a LLM component.

there are no shortcuts. you need DL.

you haven’t even said what object classes you want to consider.

As I mentioned at the beginning, I want to detect every single thing like SAM, but I want to do it without the need for classification like SAM. The subject I wanted to get ideas for this was the working logic of RPNs. I mean on what basis do they make possible object locations. If there was a situation where it was only based on color and edges, there was an idea that this could be done directly through algorithms without the need for DL. It is both costly to teach each object class by class directly to object recognition models, and even the idea that you need to train the model on all kinds of postures of this object is difficult. In summary, the only thing I’m thinking about is, is there a chance to add something to an object recognition model that cannot be recognized with the “unknown” tag but could be an object? If so, on what basis can I do this. Thank you very much.

1 Like