When using Platform deployment, intermittently, returned predictions contain false positives

Support Ultraltyics Platform
1

Hi,

I am using the Platform Deployment of a model that I have created.

(Model is custom trained based on Yolo26s-seg 8.4.163)

Intermittently, I have found that the Endpoint returns a false positive prediction.

When I check the submitted image, the positive class is not there.

I then checked the monitoring page for the Endpoint to check latest results of the saved images of what was submitted to the Endpoint for prediction.

None of images showed that a prediction was found. (correct)

Investigation:

  1. When I use the platform to to manually “drop” the image into the Endpoint to run the prediction, correctly, no prediction is found.
  2. I have used same configuration (Confidence, IOU and Image size) in both cases.
  3. When I rerun the prediction via the Endpoint, I get a false positive again.

Questions:

Should I expect the same results when using the Endpoint and manually dropping images?

Should I expect that the predictions in the saved images submitted to Endpoint be the same as the results returned from the Endpoint?

Thanks,

Steve

2

Hello.

Can you provide the code you used to run the prediction?

3 

 var client = new RestClient("https://predict-6ab663062f703b3a47ad0790-dproatj77a-km.a.run.app");
 var request = new RestRequest("", Method.Post);
 request.AddHeader("Accept", "application/json");
 request.AddHeader("Content-Type", "multipart/form-data");
 request.AddHeader("authorization", "################");
 request.AddHeader("Content-Type", "multipart/form-data");
 request.AddFile("file", FileName, "image/png");
 request.AddParameter("conf", "0.15");
 request.AddParameter("iou", "0.70");
 request.AddParameter("imgsz", "640");
                   
 var response = await client.ExecuteAsync(request);
 GetPrediction1 myDeserializedClass = JsonConvert.DeserializeObject<GetPrediction1>(response.Content);

As I am new to Yolo, I have a question please.

With regards to the parameter ‘imgsz’ does the actual image that I supply to the Endpoint have to be 640, or does the Endpoint resize it?

This is because I have found a difference in the confidence values when I submit the image via the endpoint vs when I “drop the image” on the endpoint.

Maybe, when I “drop” the image on the endpoint, the image is first resized?

But, this does not explain when I get a positive prediction, but the monitoring page of the endpoint shows the image without a prediction found.

4

Thanks for the code, Steve. Your image does not have to be 640 × 640. imgsz=640 sets the inference input size; the model’s preprocessing handles images of other dimensions. There is one important difference between your two tests: the deployment’s browser Predict tab resizes an image before uploading it, while your C# request uploads the original file. That can change confidence scores, particularly for a detection near your conf=0.15 threshold. The Platform inference documentation describes that browser-side resize. For a closer comparison, send the exact same resized image through both routes to the same deployment with the same settings, then compare their raw JSON responses.

The Monitoring gallery is not a record of every response: it captures examples on a best-effort basis, at most one per second. So a gallery image with no overlay does not, by itself, establish that it corresponds to the request that returned a positive result. If you can match a particular monitored example to a particular API response and they disagree, that warrants investigation. The Monitoring documentation explains the capture limits.

One thing to check in your snippet: the documented endpoint path ends in /predict, whereas the URL shown does not. Please confirm the actual URL being called, remove the duplicate Content-Type header and let RestSharp set the multipart boundary, and log the HTTP status and raw response.Content before deserializing it. If the discrepancy persists, please share a reproducible image, both raw responses, the request times, and the Ultralytics version reported in the responses—with your API key removed—after checking on the latest available version.