Unsupervised object detection

Question

I am trying to detect the unique/foreign objects in a conveyor. The problem is in our case is, we don't know which type of featured object is passes through conveyor along with raw material. I am familiar with object detection techniques such as yolov and detectron which can detect object based on the feature of object that we annotate. But in our case we don't know the feature of object.

I am wondering for some generic object proposal models for detection. Please give some idea is there any pre-trained unsupervised models which suits for this? or some methods or algorithm that what can i go with?. I hope i had explained my problem as much enough. Thanks in advance.

user_15098668 · Answer

I think I understood well your issue...

If you do not want to train an object detection model because you may do not have the bounding boxes corresponding to the objects, you have several options. However, I do not think there is a pretrained model that fits on your problem since you should fine-tune it, and therefore, you should have some annotations.

One think you could do, as Virgaux Pierre said, you could use some classic clustering segmentation.
On the other hand, you could use a weakly-supervised approach which it only needs image-level labels, instead of the bounding boxes. This approach could fit well if you do not need high mAP. You could use CAM, GradCAM or other techniques to obtain activation maps. Furthermore, this approaches are easy to implement with a simple NN and some forward/backward hooks.

Hope it helps.

user_11709021 · Answer

My suggestion is that you use some classical feature detection technique to detect the features of the scene. This will allow you to track object as they move.

An example is SIFT (Scale-Invariant Feature Transform). With SIFT, the keypoints vectors of translated/rotated/scaled objects are supposed to stay the same. This will allow you to track the objects of the scene as they move.

Unsupervised object detection

2 Answers