tchittesh / fovea Goto Github PK

It seems that the experiment resolution of 960x600 (0.5x) mentioned in the paper is not correctly implemented in the dataset config file, which has a resize config with (1333x800) resolution. Am I missing anything or should I believe the config file?

The paper mentions experiments on both BDD and AVHD datasets. However, I couldn't find the initial training files (.sh or .py) for BDD. Could you kindly share the BDD-related files? Thanks.

Hello, author! Thanks for your nice work.

I would like to know what unwarp_bboxes is doing. It seems like it is transforming warp bboxes back to unwarpped bboxes. However, it is unclear how you obtain the warpped bboxes from original bboxes. Could you share the code to do that?

Greetings! Is there any chance that you'll make the training script publicly available any time soon?

Hi dear authors!

Thank you very much for sharing your work! I am quite interested in the concept of intelligent down-scaling but I am struggling with getting a modular version of this particular block to run :

My code keep running into dimension mismatch error while calculating weighted offsets on line 169 of the same file. And installing mmdetection is proving to be hassle so I could not get your whole code to run to check it myself.

So lets say I have an RGB image of size 1024x1024x3 and a saliency image of 1024x1024, what would be the values of the following parameters: output_shape and grid_shape ? (Because padding and other parameters are derived from it and the problem comes with matrix multiplication compatibility)

I have tried looking at the base paper but the math was not too clear to me. Apologies for my naivety! Any help will be much appreciated!!