search

MarvinTeichmann / KittiSeg

A Kitti Road Segmentation model implemented in tensorflow.
MIT License
911 stars 403 forks source link

Extremely bad results and recurring exception #187

Open tombenj opened 6 years ago

tombenj commented 6 years ago

Results on the pre-trained data are turning really awful: car-road_green

And the recurring exception that everyone seems to get over and over again is: InvalidArgumentError (see above for traceback): Number of ways to split should evenly divide the split dimension, but got split_dim 3 (size = 4) and num_split 3

@MarvinTeichmann is anyone able to help?

athityakumar commented 6 years ago

Hey all.

The above exception is being caused due to the presence of alpha channel in the input image, which can be removed by specifying this:

# Existing line in demo.py
image = scp.misc.imread(input_image)

# Add this line next to it
image = image[:,:,:3]
athityakumar commented 6 years ago

I even tried increasing the brightness as discussed in #129 , but that doesn't seem to work. Is there any way the input image needs to be pre-processed/transformed that can improve the results for road detection?

Ping @MarvinTeichmann