evaluator.airborne_detection.TimeoutException: Prediction timed out!

Hi, @shivam @zontakm9

I wanna to run my tracker. And INFERENCE_PER_FLIGHT_TIMEOUT_SECONDS is setting to 600 for each flight_id (about 1200 imgs). It is easy to meet TimeoutException for our model.

Due to the high resolution of image and small object, I believe we need to extend INFERENCE_PER_FLIGHT_TIMEOUT_SECONDS.

Could you provided some suggestion? Thanks!

Hi @octo,

500ms/frame is a hard timeout and unfortunately can’t be changed.
We will still discuss in case it can be relaxed further, and let you know. :sweat_smile:

You can potentially try batching during inference in case your codebase isn’t using resources fully to fasten it up.

1 Like

Please suggest how much time you think would be sufficient for your model inference?
One suggestion that I can give you to accelerate your performance is to estimate prior based on training set where the objects typically appear and try to process only this part of the image
Do not forget to provide the final results in the original coordinate system

I know the timeout has been increased, but for the model/approaches to be useful, worth keeping in mind it’s supposed to run on a drone with a much less powerful GPU in real-time with a time budget of only 100ms per frame.