I'm trying to reproduce the article's results using the IOG_PASCAL_SBD.pth, but the results are very different. For example, on the Rooftop dataset without fine-tuning, I'm scoring 45.10, while the article reports above 90 IoU.
I'm not sure what I'm doing wrong. Is any pre-processing necessary?
Currently, I'm using the cropped BRG image from OpenCV as input and computing the Gaussians using the default parameters of the make_gaussian function.
Hi, @shiyinzhang
I'm trying to reproduce the article's results using the IOG_PASCAL_SBD.pth, but the results are very different. For example, on the Rooftop dataset without fine-tuning, I'm scoring 45.10, while the article reports above 90 IoU.
I'm not sure what I'm doing wrong. Is any pre-processing necessary? Currently, I'm using the cropped BRG image from OpenCV as input and computing the Gaussians using the default parameters of the
make_gaussian
function.Thank you,