It makes sense to update the model in tests too.

Hi @asmorkalov, V1 and V2 have similar structures. And totally the same pre-process and post-process. I'm not sure if it is necessary.

Performance improvement says opposite. It's not just a retrained model. There are some changes in architecture or at least channels.

The final decision, we keep both Nanotrack v1 and v2 models at regression test. Since the model is quite lightweight, it will not bring too much burden to CI.

@zihaomu please squash the commits. Waiting for OpenCV Extra PR approve.

BTW, it makes sense to add performance tests of dnn for these (and other) models: https://github.com/opencv/opencv/blob/4.x/modules/dnn/perf/perf_net.cpp

(the last model was added there in 2020)

Hi @alalek, the tracking model is different. It consists of two models, the first model for extracting first frame features, and the second model for extracting and locating the object.
The first mode expected input is an image, and the second model expected two muti-dimension Mat as inputs. If we want to add the performance test, we need to test these two models at single performance test.

BTW, because Nanotrack V2 is superior to V1 in both speed and accuracy a lot, we do not expect users to choose between using Nanotrack V1 or V2. We only want users to use the V2 version of tracking.

@zihaomu, support nanotrackv3?
Somthing wrong when i change model to nanotrackv3.

How to deal? thank you

Hi @ycxia, v3 has some changes, but not much. Will try to add nanotrack v3 next week.