RangiLyu / nanodet

NanoDet-Plusāš”Super fast and lightweight anchor-free object detection model. šŸ”„Only 980 KB(int8) / 1.8MB (fp16) and run 97FPS on cellphonešŸ”„
Apache License 2.0
5.77k stars 1.04k forks source link

Dynamic input size in NanoDet #562

Open Borzyszkowski opened 7 months ago

Borzyszkowski commented 7 months ago

Hello,

I'm wondering how dynamic input size is supported in Nanodet.

I noticed that there's no padding in the input image as long as width and height are divisible to 32. I train the model with shape 192x192, but during inference I can use it with various input sizes.

Adaptive pooling layers are typically used to support dynamic input size, but I didn't see them in the backbone. How is it handled in Nanodet?