Open XiaolongQian opened 3 months ago
It is due to the normalization part in the decoder (F.layer_norm
) that is in the log space, while the scale and shit are re-computed in the linear space in the V2, we will try to clean the model and remove this normalization, but it requires a re-training and we are working on it. This happens for close-by scenes, where the numbers are snapped to zero.
Thank you for your efficient and timely response. I would be grateful if you could make the retrained model available!
Thanks for this outstanding work! When I use UnidepthV2 to infer a 1280x1920 RGB image, why there are always some missing values in the lower left and lower right corners of the predicted depth map? The input image and predicted depth map are shown in the figure below. There are missing values in the lower right corner of the predicted depth map.
The inference code I used is from Get started. I would be grateful if you could tell me the reason for this missing value.