How to avoid the object detector to focus on some objects

vision
1

Hello everyone,

I’m working with the Faster RCNN to predict the bounding boxes of some particular objects inside a fixed environment. Basically what I want to predict are out-of-place bolts inside a scene. I already trained my model with synthetic images and validating it on real images and fortunately I get some discreet results. The detector is quite confident on locating the bolts (scores > 0.98) but it gets confused with objects with similar shape. An example could be the rivets.

I created my dataset including only images with objects and their bounding boxes but is there a way to actually “suggest” the model even to not consider some objects instead of telling it to locate those bolts?
Thanks in advance

2

You could add these negative examples to the dataset and don’t provide valid bounding boxes for all negative instances, which would train the model to ignore them.

1 Like
3

Do you mean by providing empty bounding boxes? Suppose there is an object that I don’t want to be recognized, i can simply annotate the image with

target = {}
target["boxes"] = torch.zeros((0, 4), dtype=torch.float32)
target["labels"] = torch.zeros((0, 1), dtype=torch.int64)

… or maybe by considering the whole image as a bounding box

4

I don’t know if it would be necessary to provide empty boxes or just not annotate the unwanted objects.

5

Hi! Thanks for your advice. I think images without annotations would be ignored at all during loading. I’m not labeling the unwanted objects. How to provide images with the unwanted objects only and not give annotations?