search

IDEA-Research / DINO

[ICLR 2023] Official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection"
Apache License 2.0
2.19k stars 243 forks source link

Prediction of 4K images using default parameters and weights from COCO dataset #222

Open pascual-tejero opened 1 year ago

pascual-tejero commented 1 year ago

I have a question about your model. I want to predict images with a resolution 3840x2160 using the model DINO-4scale with SwinL as backbone for a object detection task. I used the weights (_checkpoint0029_4scaleswin.pth) you uploaded from COCO dataset.

However, I am getting really poor results, almost there is no good prediction with that image resolution. Why is that? Shouldn't be the model scalable in image size? Or is it because the model is trained with an image size of 2000 pixels at most and I have to train the model for larger image resolutions?

hdmjdp commented 3 months ago

have you solved this problem？

pascual-tejero commented 3 months ago

Not yet