Problem about different tflite model's invoke time cost? about tflite-support HOT 4 CLOSED

@LeopoldACC You can try the following ways to optimize the performance.
(1) Use a model with smaller input size. The one you are using is 512x1024, which is pretty large.
(2) Config Interpreter with multithread, see the example here
(3) Use GPU delegate. See the instruction here

Hope it helps.

from tflite-support.

In my task, the segmentation object is slender(It's one or two pixels wide at 1920 * 1080 resolution). If I continue down sampling, the information of this object may be lost, and the accuracy will inevitably decline seriously. If I cut 256 * 256 blocks from 1024 * 512, I need to execute invoke () 8 times for an image. This task is really difficult for me to achieve real-time and high precision.But thank you a lot for your timely feedback!

from tflite-support.

Usually the segmentation model will be applied on a smaller size of image such as 316x316 and then upscale the output mask to the original image size. In that way, you'll get a much faster inference.

from tflite-support.

Closing this issue. Feel free to reopen it if you have other questions.

from tflite-support.