Closed lakshjaisinghani closed 3 years ago
Hi, it seems to you didn't resize the image and gave the network a full resolution. You need to resize to 416x128 to use the pretrained network.
the pose network has been trained with 5 sequence images, because that is the training that got the best results for disparity estimation. The 3-frames version also works, but not as good.
I see, Thanks a ton!
Hey Clement, Thanks for the great code, but I'm wondering what I may be doing wrong.
Link to ipynb (https://github.com/Monash-Connected-Autonomous-Vehicle/UMDE-study/blob/main/visualizer.ipynb)
When I run a single image through the disparity network (using _dispnet_model_best.pth.tar_ as the weights). I get something like this (1/output):
The image had been tested with and without normalization before the forward pass, but the result is always the same. I'm wondering what's wrong?
I was also wondering why is the trained pose network requiring 5 sequence images?