Using Computer Vision to Classify Mixed Waste Images

I’m exploring a small computer-vision use case and would be interested in hearing how others would approach it on AMD hardware.

The basic idea is to let a user upload a photo containing several unwanted items and identify what is actually visible in the image. A single photo could contain things like furniture, appliances, mattresses, boxes, electronics, or yard debris.

I think object detection would make more sense than simple image classification here because there can be several different objects in the same image. The output could be something simple like:

  • furniture
  • appliances
  • electronics
  • mattresses
  • yard debris
  • general household items
  • unknown

One part I’m working through is how to define the categories before training or prompting the model. I’m using what junk removal companies take as an example of how real-world service categories can be defined before building the classification layer.

I don’t want the model to invent categories every time it sees something unfamiliar. I’d rather give it a controlled set of categories and let it return unknown or needs_review when the image doesn’t provide enough confidence.

Mixed images are another challenge. If someone uploads one photo containing a sofa, an appliance and several boxes, the system should ideally identify each object separately instead of assigning one label to the entire image.

I’m also curious about the practical side of running this kind of workload. The application doesn’t need video-rate inference, but the result should come back quickly enough that a user isn’t sitting around waiting for an image analysis to finish.

For anyone who has worked with AMD hardware and computer-vision inference:

  1. Would you use object detection or segmentation for this type of image?
  2. Which AMD software stack would you recommend for running inference?
  3. Would you optimize the model for edge inference if the workload was relatively small?
  4. How would you handle multiple objects with different confidence levels in one image?
  5. Would you keep an explicit unknown class, or handle uncertainty outside the model?

I’m mainly interested in practical implementation experience. I’m trying to understand where the real trade-offs are between model accuracy, inference speed, hardware requirements, and the amount of preprocessing needed.

1 Like

One thing I’m thinking about testing first is the mixed-image case with a small controlled set rather than starting with the full workload.

For example, a photo with 3–4 clearly different objects and an explicit unknown class. That should give me a better way to compare object detection vs segmentation and see how the model behaves when one item doesn’t match the expected categories.

If anyone has done something similar on AMD edge hardware, I’d especially like to know whether you found the inference latency or model accuracy to be the bigger constraint in practice.

Hello, I read the page you linked. Its categories are not visual ones. The split there is accepted, special handling, not accepted, and what decides it is mostly invisible in a photo. A washer and a fridge look nearly the same, but the fridge triggers refrigerant recovery under EPA Section 608 and a fee. Wet paint and dried paint sit in the same can. So I would have the detector predict plain object nouns, sofa, fridge, mattress, box, yard debris, and put the accepted or special handling call in a lookup table on top. Unknown and needs_review then live in that layer plus a confidence threshold, rather than as a trained class. It also means you edit a table when the disposal rules change instead of retraining.

That’s a really good point. I was thinking mainly about identifying the objects first, but you’re right that the disposal category can depend on things that aren’t visible in the image.

In that case, would you recommend keeping the vision model limited to object identification and then using a separate rules/lookup layer for things like special handling, accepted items, and unknown cases? That seems safer than having the model make the final disposal decision itself.

Yes do not let the model make the final decision just for now though. In the process, you will see how it is doing and you will understand on which areas to work on, so later it can be fully automated.

That makes sense. I’d probably keep the first version intentionally conservative and log the model’s confidence for each detected object along with the rule-layer outcome. That would make it easier to see whether the weak point is the visual identification itself or the mapping from an identified object to a disposal rule. For mixed images, I’d also track the cases where multiple objects are detected but one falls below the confidence threshold. That kind of error log should give you a pretty clear roadmap for what needs improving before turning any part of the process into full automation.

1 Like