Object Detection vs Instance Segmentation for anomaly detection, which one should I choose?

🤝 Community Help
1

My Question: Should I use Object Detection or Instance Segmentation for my use case?

I’m working on a hospital CCTV anomaly detection project. My classes are:

  • guard (blue uniform)
  • person (visitors)
  • child
  • person_with_paper (normal)
  • person_without_paper (anomaly)

The goal is to detect when a person is NOT carrying their document/slip — that is the anomaly.

  • Project Type: Object Detection (but unsure if I should switch to Segmentation)

  • Operating System & Browser: Windows / Chrome

  • Project Universe Link or Workspace/Project ID: workspace: mks-workspace-iihjl and ID: 2_stage_doc

  • Do you grant Roboflow Support permission to access your Workspace for troubleshooting? Yes

    image
    image1865×933 186 KB

2

Hey! My immediate thought is you do want Object Detection so nice work on that choice. And then I would simplify things and just find a guard, person, child initially. Then crop that detection and look for a paper using a separate classification model that you train on cropped images.

But that being said, I have lots of questions! :slight_smile: It looks like maybe a guard does not need a paper. If so, then your first model trains on guard vs visitor and when you go to crop detections you only worry about visitors.

It also seems likely a paper could be in view for a few frames and then out of view and then back in view. If you only expect one person in a frame at a time that would not be a big deal. But if you have a few people like in your example, you might want to throw on a tracking feature for the person to make sure that “at some point” in the video they have a paper.

As you build out pieces and start to hit other roadblocks continue to drop your questions in here!