How Triton inference server always compare the current frame infer result with the previous one

[Komoro2023]

Hi there, I want to reduce the false alarms of fire and smoke detection model, and our models are deployed on triton inference server, so I try to find if there is a way that on triton inference server it always compare the current infer result with previous infere result, if the area of bbox of detection is changing, so it is probably a real alarm, and give the response, or it is a false alarm, give a empty repsonse. so please help me find a way or is there a feature can solve my problem?
Many thanks.

I found Triton Response Cache feature probably could do this work, but don't know how to get the previous infer result from response cache, and there is no any example for extracting from response cache

[tanmayv25]

@Komoro2023 You can add a custom python backend model in front of your fire and smoke detection model building pipeline either using ensemble or BLS. See here for more information.

The custom model can store the last result and compare it with the incoming result. If they are identical, it can return empty response, otherwise it can return the needful.

Response Cache is placed in front of the model. If the input to the model is identical to the previously cached input then response data is pulled from the cache instead of executing inference on model.