Run TAPIR Inference without resizing the video

[vrk7]

Hi tEAM,

Can the inference be done without resizing the video frames?

[cdoersch]

Yes, TAPIR can be run on any resolution; internally it will perform its initialization on 256x256 and then refine on progressively higher resolution. It should work out-of-the-box in our colab demo.

[vrk7]

My issue is, i have annotation of segments in COCO format. When i converted this into 2D points, should I have the points according to the original frame or reduced frame?

[cdoersch]

They should be scaled relative to the size of the video that you input into TAPIR. I.e. if its 512x512, your points should be in the range [0,512].